1 //===-- AutoUpgrade.cpp - Implement auto-upgrade helper functions ---------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the auto-upgrade helper functions.
11 // This is where deprecated IR intrinsics and other IR features are updated to
12 // current specifications.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "llvm/IR/AutoUpgrade.h"
17 #include "llvm/ADT/StringSwitch.h"
18 #include "llvm/IR/CFG.h"
19 #include "llvm/IR/CallSite.h"
20 #include "llvm/IR/Constants.h"
21 #include "llvm/IR/DIBuilder.h"
22 #include "llvm/IR/DebugInfo.h"
23 #include "llvm/IR/DiagnosticInfo.h"
24 #include "llvm/IR/Function.h"
25 #include "llvm/IR/IRBuilder.h"
26 #include "llvm/IR/Instruction.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/Module.h"
30 #include "llvm/Support/ErrorHandling.h"
31 #include "llvm/Support/Regex.h"
32 #include <cstring>
33 using namespace llvm;
34 
35 static void rename(GlobalValue *GV) { GV->setName(GV->getName() + ".old"); }
36 
37 // Upgrade the declarations of the SSE4.1 ptest intrinsics whose arguments have
38 // changed their type from v4f32 to v2i64.
39 static bool UpgradePTESTIntrinsic(Function* F, Intrinsic::ID IID,
40                                   Function *&NewFn) {
41   // Check whether this is an old version of the function, which received
42   // v4f32 arguments.
43   Type *Arg0Type = F->getFunctionType()->getParamType(0);
44   if (Arg0Type != VectorType::get(Type::getFloatTy(F->getContext()), 4))
45     return false;
46 
47   // Yes, it's old, replace it with new version.
48   rename(F);
49   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
50   return true;
51 }
52 
53 // Upgrade the declarations of intrinsic functions whose 8-bit immediate mask
54 // arguments have changed their type from i32 to i8.
55 static bool UpgradeX86IntrinsicsWith8BitMask(Function *F, Intrinsic::ID IID,
56                                              Function *&NewFn) {
57   // Check that the last argument is an i32.
58   Type *LastArgType = F->getFunctionType()->getParamType(
59      F->getFunctionType()->getNumParams() - 1);
60   if (!LastArgType->isIntegerTy(32))
61     return false;
62 
63   // Move this function aside and map down.
64   rename(F);
65   NewFn = Intrinsic::getDeclaration(F->getParent(), IID);
66   return true;
67 }
68 
69 static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
70   // All of the intrinsics matches below should be marked with which llvm
71   // version started autoupgrading them. At some point in the future we would
72   // like to use this information to remove upgrade code for some older
73   // intrinsics. It is currently undecided how we will determine that future
74   // point.
75   if (Name=="ssse3.pabs.b.128" || // Added in 6.0
76       Name=="ssse3.pabs.w.128" || // Added in 6.0
77       Name=="ssse3.pabs.d.128" || // Added in 6.0
78       Name.startswith("avx2.pabs.") || // Added in 6.0
79       Name.startswith("avx512.mask.pabs.") || // Added in 6.0
80       Name.startswith("avx512.mask.pbroadcast") || // Added in 6.0
81       Name.startswith("sse2.pcmpeq.") || // Added in 3.1
82       Name.startswith("sse2.pcmpgt.") || // Added in 3.1
83       Name.startswith("avx2.pcmpeq.") || // Added in 3.1
84       Name.startswith("avx2.pcmpgt.") || // Added in 3.1
85       Name.startswith("avx512.mask.pcmpeq.") || // Added in 3.9
86       Name.startswith("avx512.mask.pcmpgt.") || // Added in 3.9
87       Name.startswith("avx.vperm2f128.") || // Added in 6.0
88       Name == "avx2.vperm2i128" || // Added in 6.0
89       Name == "sse.add.ss" || // Added in 4.0
90       Name == "sse2.add.sd" || // Added in 4.0
91       Name == "sse.sub.ss" || // Added in 4.0
92       Name == "sse2.sub.sd" || // Added in 4.0
93       Name == "sse.mul.ss" || // Added in 4.0
94       Name == "sse2.mul.sd" || // Added in 4.0
95       Name == "sse.div.ss" || // Added in 4.0
96       Name == "sse2.div.sd" || // Added in 4.0
97       Name == "sse41.pmaxsb" || // Added in 3.9
98       Name == "sse2.pmaxs.w" || // Added in 3.9
99       Name == "sse41.pmaxsd" || // Added in 3.9
100       Name == "sse2.pmaxu.b" || // Added in 3.9
101       Name == "sse41.pmaxuw" || // Added in 3.9
102       Name == "sse41.pmaxud" || // Added in 3.9
103       Name == "sse41.pminsb" || // Added in 3.9
104       Name == "sse2.pmins.w" || // Added in 3.9
105       Name == "sse41.pminsd" || // Added in 3.9
106       Name == "sse2.pminu.b" || // Added in 3.9
107       Name == "sse41.pminuw" || // Added in 3.9
108       Name == "sse41.pminud" || // Added in 3.9
109       Name.startswith("avx512.mask.pshuf.b.") || // Added in 4.0
110       Name.startswith("avx2.pmax") || // Added in 3.9
111       Name.startswith("avx2.pmin") || // Added in 3.9
112       Name.startswith("avx512.mask.pmax") || // Added in 4.0
113       Name.startswith("avx512.mask.pmin") || // Added in 4.0
114       Name.startswith("avx2.vbroadcast") || // Added in 3.8
115       Name.startswith("avx2.pbroadcast") || // Added in 3.8
116       Name.startswith("avx.vpermil.") || // Added in 3.1
117       Name.startswith("sse2.pshuf") || // Added in 3.9
118       Name.startswith("avx512.pbroadcast") || // Added in 3.9
119       Name.startswith("avx512.mask.broadcast.s") || // Added in 3.9
120       Name.startswith("avx512.mask.movddup") || // Added in 3.9
121       Name.startswith("avx512.mask.movshdup") || // Added in 3.9
122       Name.startswith("avx512.mask.movsldup") || // Added in 3.9
123       Name.startswith("avx512.mask.pshuf.d.") || // Added in 3.9
124       Name.startswith("avx512.mask.pshufl.w.") || // Added in 3.9
125       Name.startswith("avx512.mask.pshufh.w.") || // Added in 3.9
126       Name.startswith("avx512.mask.shuf.p") || // Added in 4.0
127       Name.startswith("avx512.mask.vpermil.p") || // Added in 3.9
128       Name.startswith("avx512.mask.perm.df.") || // Added in 3.9
129       Name.startswith("avx512.mask.perm.di.") || // Added in 3.9
130       Name.startswith("avx512.mask.punpckl") || // Added in 3.9
131       Name.startswith("avx512.mask.punpckh") || // Added in 3.9
132       Name.startswith("avx512.mask.unpckl.") || // Added in 3.9
133       Name.startswith("avx512.mask.unpckh.") || // Added in 3.9
134       Name.startswith("avx512.mask.pand.") || // Added in 3.9
135       Name.startswith("avx512.mask.pandn.") || // Added in 3.9
136       Name.startswith("avx512.mask.por.") || // Added in 3.9
137       Name.startswith("avx512.mask.pxor.") || // Added in 3.9
138       Name.startswith("avx512.mask.and.") || // Added in 3.9
139       Name.startswith("avx512.mask.andn.") || // Added in 3.9
140       Name.startswith("avx512.mask.or.") || // Added in 3.9
141       Name.startswith("avx512.mask.xor.") || // Added in 3.9
142       Name.startswith("avx512.mask.padd.") || // Added in 4.0
143       Name.startswith("avx512.mask.psub.") || // Added in 4.0
144       Name.startswith("avx512.mask.pmull.") || // Added in 4.0
145       Name.startswith("avx512.mask.cvtdq2pd.") || // Added in 4.0
146       Name.startswith("avx512.mask.cvtudq2pd.") || // Added in 4.0
147       Name.startswith("avx512.mask.pmul.dq.") || // Added in 4.0
148       Name.startswith("avx512.mask.pmulu.dq.") || // Added in 4.0
149       Name.startswith("avx512.mask.packsswb.") || // Added in 5.0
150       Name.startswith("avx512.mask.packssdw.") || // Added in 5.0
151       Name.startswith("avx512.mask.packuswb.") || // Added in 5.0
152       Name.startswith("avx512.mask.packusdw.") || // Added in 5.0
153       Name.startswith("avx512.mask.cmp.b") || // Added in 5.0
154       Name.startswith("avx512.mask.cmp.d") || // Added in 5.0
155       Name.startswith("avx512.mask.cmp.q") || // Added in 5.0
156       Name.startswith("avx512.mask.cmp.w") || // Added in 5.0
157       Name.startswith("avx512.mask.ucmp.") || // Added in 5.0
158       Name == "avx512.mask.add.pd.128" || // Added in 4.0
159       Name == "avx512.mask.add.pd.256" || // Added in 4.0
160       Name == "avx512.mask.add.ps.128" || // Added in 4.0
161       Name == "avx512.mask.add.ps.256" || // Added in 4.0
162       Name == "avx512.mask.div.pd.128" || // Added in 4.0
163       Name == "avx512.mask.div.pd.256" || // Added in 4.0
164       Name == "avx512.mask.div.ps.128" || // Added in 4.0
165       Name == "avx512.mask.div.ps.256" || // Added in 4.0
166       Name == "avx512.mask.mul.pd.128" || // Added in 4.0
167       Name == "avx512.mask.mul.pd.256" || // Added in 4.0
168       Name == "avx512.mask.mul.ps.128" || // Added in 4.0
169       Name == "avx512.mask.mul.ps.256" || // Added in 4.0
170       Name == "avx512.mask.sub.pd.128" || // Added in 4.0
171       Name == "avx512.mask.sub.pd.256" || // Added in 4.0
172       Name == "avx512.mask.sub.ps.128" || // Added in 4.0
173       Name == "avx512.mask.sub.ps.256" || // Added in 4.0
174       Name == "avx512.mask.max.pd.128" || // Added in 5.0
175       Name == "avx512.mask.max.pd.256" || // Added in 5.0
176       Name == "avx512.mask.max.ps.128" || // Added in 5.0
177       Name == "avx512.mask.max.ps.256" || // Added in 5.0
178       Name == "avx512.mask.min.pd.128" || // Added in 5.0
179       Name == "avx512.mask.min.pd.256" || // Added in 5.0
180       Name == "avx512.mask.min.ps.128" || // Added in 5.0
181       Name == "avx512.mask.min.ps.256" || // Added in 5.0
182       Name.startswith("avx512.mask.vpermilvar.") || // Added in 4.0
183       Name.startswith("avx512.mask.psll.d") || // Added in 4.0
184       Name.startswith("avx512.mask.psll.q") || // Added in 4.0
185       Name.startswith("avx512.mask.psll.w") || // Added in 4.0
186       Name.startswith("avx512.mask.psra.d") || // Added in 4.0
187       Name.startswith("avx512.mask.psra.q") || // Added in 4.0
188       Name.startswith("avx512.mask.psra.w") || // Added in 4.0
189       Name.startswith("avx512.mask.psrl.d") || // Added in 4.0
190       Name.startswith("avx512.mask.psrl.q") || // Added in 4.0
191       Name.startswith("avx512.mask.psrl.w") || // Added in 4.0
192       Name.startswith("avx512.mask.pslli") || // Added in 4.0
193       Name.startswith("avx512.mask.psrai") || // Added in 4.0
194       Name.startswith("avx512.mask.psrli") || // Added in 4.0
195       Name.startswith("avx512.mask.psllv") || // Added in 4.0
196       Name.startswith("avx512.mask.psrav") || // Added in 4.0
197       Name.startswith("avx512.mask.psrlv") || // Added in 4.0
198       Name.startswith("sse41.pmovsx") || // Added in 3.8
199       Name.startswith("sse41.pmovzx") || // Added in 3.9
200       Name.startswith("avx2.pmovsx") || // Added in 3.9
201       Name.startswith("avx2.pmovzx") || // Added in 3.9
202       Name.startswith("avx512.mask.pmovsx") || // Added in 4.0
203       Name.startswith("avx512.mask.pmovzx") || // Added in 4.0
204       Name.startswith("avx512.mask.lzcnt.") || // Added in 5.0
205       Name == "sse2.cvtdq2pd" || // Added in 3.9
206       Name == "sse2.cvtps2pd" || // Added in 3.9
207       Name == "avx.cvtdq2.pd.256" || // Added in 3.9
208       Name == "avx.cvt.ps2.pd.256" || // Added in 3.9
209       Name.startswith("avx.vinsertf128.") || // Added in 3.7
210       Name == "avx2.vinserti128" || // Added in 3.7
211       Name.startswith("avx512.mask.insert") || // Added in 4.0
212       Name.startswith("avx.vextractf128.") || // Added in 3.7
213       Name == "avx2.vextracti128" || // Added in 3.7
214       Name.startswith("avx512.mask.vextract") || // Added in 4.0
215       Name.startswith("sse4a.movnt.") || // Added in 3.9
216       Name.startswith("avx.movnt.") || // Added in 3.2
217       Name.startswith("avx512.storent.") || // Added in 3.9
218       Name == "sse41.movntdqa" || // Added in 5.0
219       Name == "avx2.movntdqa" || // Added in 5.0
220       Name == "avx512.movntdqa" || // Added in 5.0
221       Name == "sse2.storel.dq" || // Added in 3.9
222       Name.startswith("sse.storeu.") || // Added in 3.9
223       Name.startswith("sse2.storeu.") || // Added in 3.9
224       Name.startswith("avx.storeu.") || // Added in 3.9
225       Name.startswith("avx512.mask.storeu.") || // Added in 3.9
226       Name.startswith("avx512.mask.store.p") || // Added in 3.9
227       Name.startswith("avx512.mask.store.b.") || // Added in 3.9
228       Name.startswith("avx512.mask.store.w.") || // Added in 3.9
229       Name.startswith("avx512.mask.store.d.") || // Added in 3.9
230       Name.startswith("avx512.mask.store.q.") || // Added in 3.9
231       Name.startswith("avx512.mask.loadu.") || // Added in 3.9
232       Name.startswith("avx512.mask.load.") || // Added in 3.9
233       Name == "sse42.crc32.64.8" || // Added in 3.4
234       Name.startswith("avx.vbroadcast.s") || // Added in 3.5
235       Name.startswith("avx512.mask.palignr.") || // Added in 3.9
236       Name.startswith("avx512.mask.valign.") || // Added in 4.0
237       Name.startswith("sse2.psll.dq") || // Added in 3.7
238       Name.startswith("sse2.psrl.dq") || // Added in 3.7
239       Name.startswith("avx2.psll.dq") || // Added in 3.7
240       Name.startswith("avx2.psrl.dq") || // Added in 3.7
241       Name.startswith("avx512.psll.dq") || // Added in 3.9
242       Name.startswith("avx512.psrl.dq") || // Added in 3.9
243       Name == "sse41.pblendw" || // Added in 3.7
244       Name.startswith("sse41.blendp") || // Added in 3.7
245       Name.startswith("avx.blend.p") || // Added in 3.7
246       Name == "avx2.pblendw" || // Added in 3.7
247       Name.startswith("avx2.pblendd.") || // Added in 3.7
248       Name.startswith("avx.vbroadcastf128") || // Added in 4.0
249       Name == "avx2.vbroadcasti128" || // Added in 3.7
250       Name.startswith("avx512.mask.broadcastf32x4.") || // Added in 6.0
251       Name.startswith("avx512.mask.broadcastf64x2.") || // Added in 6.0
252       Name.startswith("avx512.mask.broadcasti32x4.") || // Added in 6.0
253       Name.startswith("avx512.mask.broadcasti64x2.") || // Added in 6.0
254       Name == "avx512.mask.broadcastf32x8.512" || // Added in 6.0
255       Name == "avx512.mask.broadcasti32x8.512" || // Added in 6.0
256       Name == "avx512.mask.broadcastf64x4.512" || // Added in 6.0
257       Name == "avx512.mask.broadcasti64x4.512" || // Added in 6.0
258       Name == "xop.vpcmov" || // Added in 3.8
259       Name == "xop.vpcmov.256" || // Added in 5.0
260       Name.startswith("avx512.mask.move.s") || // Added in 4.0
261       Name.startswith("avx512.cvtmask2") || // Added in 5.0
262       (Name.startswith("xop.vpcom") && // Added in 3.2
263        F->arg_size() == 2) ||
264       Name.startswith("sse2.pavg") || // Added in 6.0
265       Name.startswith("avx2.pavg") || // Added in 6.0
266       Name.startswith("avx512.mask.pavg")) // Added in 6.0
267     return true;
268 
269   return false;
270 }
271 
272 static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
273                                         Function *&NewFn) {
274   // Only handle intrinsics that start with "x86.".
275   if (!Name.startswith("x86."))
276     return false;
277   // Remove "x86." prefix.
278   Name = Name.substr(4);
279 
280   if (ShouldUpgradeX86Intrinsic(F, Name)) {
281     NewFn = nullptr;
282     return true;
283   }
284 
285   // SSE4.1 ptest functions may have an old signature.
286   if (Name.startswith("sse41.ptest")) { // Added in 3.2
287     if (Name.substr(11) == "c")
288       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestc, NewFn);
289     if (Name.substr(11) == "z")
290       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestz, NewFn);
291     if (Name.substr(11) == "nzc")
292       return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
293   }
294   // Several blend and other instructions with masks used the wrong number of
295   // bits.
296   if (Name == "sse41.insertps") // Added in 3.6
297     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
298                                             NewFn);
299   if (Name == "sse41.dppd") // Added in 3.6
300     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
301                                             NewFn);
302   if (Name == "sse41.dpps") // Added in 3.6
303     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
304                                             NewFn);
305   if (Name == "sse41.mpsadbw") // Added in 3.6
306     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
307                                             NewFn);
308   if (Name == "avx.dp.ps.256") // Added in 3.6
309     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
310                                             NewFn);
311   if (Name == "avx2.mpsadbw") // Added in 3.6
312     return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
313                                             NewFn);
314 
315   // frcz.ss/sd may need to have an argument dropped. Added in 3.2
316   if (Name.startswith("xop.vfrcz.ss") && F->arg_size() == 2) {
317     rename(F);
318     NewFn = Intrinsic::getDeclaration(F->getParent(),
319                                       Intrinsic::x86_xop_vfrcz_ss);
320     return true;
321   }
322   if (Name.startswith("xop.vfrcz.sd") && F->arg_size() == 2) {
323     rename(F);
324     NewFn = Intrinsic::getDeclaration(F->getParent(),
325                                       Intrinsic::x86_xop_vfrcz_sd);
326     return true;
327   }
328   // Upgrade any XOP PERMIL2 index operand still using a float/double vector.
329   if (Name.startswith("xop.vpermil2")) { // Added in 3.9
330     auto Idx = F->getFunctionType()->getParamType(2);
331     if (Idx->isFPOrFPVectorTy()) {
332       rename(F);
333       unsigned IdxSize = Idx->getPrimitiveSizeInBits();
334       unsigned EltSize = Idx->getScalarSizeInBits();
335       Intrinsic::ID Permil2ID;
336       if (EltSize == 64 && IdxSize == 128)
337         Permil2ID = Intrinsic::x86_xop_vpermil2pd;
338       else if (EltSize == 32 && IdxSize == 128)
339         Permil2ID = Intrinsic::x86_xop_vpermil2ps;
340       else if (EltSize == 64 && IdxSize == 256)
341         Permil2ID = Intrinsic::x86_xop_vpermil2pd_256;
342       else
343         Permil2ID = Intrinsic::x86_xop_vpermil2ps_256;
344       NewFn = Intrinsic::getDeclaration(F->getParent(), Permil2ID);
345       return true;
346     }
347   }
348 
349   return false;
350 }
351 
352 static bool UpgradeIntrinsicFunction1(Function *F, Function *&NewFn) {
353   assert(F && "Illegal to upgrade a non-existent Function.");
354 
355   // Quickly eliminate it, if it's not a candidate.
356   StringRef Name = F->getName();
357   if (Name.size() <= 8 || !Name.startswith("llvm."))
358     return false;
359   Name = Name.substr(5); // Strip off "llvm."
360 
361   switch (Name[0]) {
362   default: break;
363   case 'a': {
364     if (Name.startswith("arm.rbit") || Name.startswith("aarch64.rbit")) {
365       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::bitreverse,
366                                         F->arg_begin()->getType());
367       return true;
368     }
369     if (Name.startswith("arm.neon.vclz")) {
370       Type* args[2] = {
371         F->arg_begin()->getType(),
372         Type::getInt1Ty(F->getContext())
373       };
374       // Can't use Intrinsic::getDeclaration here as it adds a ".i1" to
375       // the end of the name. Change name from llvm.arm.neon.vclz.* to
376       //  llvm.ctlz.*
377       FunctionType* fType = FunctionType::get(F->getReturnType(), args, false);
378       NewFn = Function::Create(fType, F->getLinkage(),
379                                "llvm.ctlz." + Name.substr(14), F->getParent());
380       return true;
381     }
382     if (Name.startswith("arm.neon.vcnt")) {
383       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
384                                         F->arg_begin()->getType());
385       return true;
386     }
387     Regex vldRegex("^arm\\.neon\\.vld([1234]|[234]lane)\\.v[a-z0-9]*$");
388     if (vldRegex.match(Name)) {
389       auto fArgs = F->getFunctionType()->params();
390       SmallVector<Type *, 4> Tys(fArgs.begin(), fArgs.end());
391       // Can't use Intrinsic::getDeclaration here as the return types might
392       // then only be structurally equal.
393       FunctionType* fType = FunctionType::get(F->getReturnType(), Tys, false);
394       NewFn = Function::Create(fType, F->getLinkage(),
395                                "llvm." + Name + ".p0i8", F->getParent());
396       return true;
397     }
398     Regex vstRegex("^arm\\.neon\\.vst([1234]|[234]lane)\\.v[a-z0-9]*$");
399     if (vstRegex.match(Name)) {
400       static const Intrinsic::ID StoreInts[] = {Intrinsic::arm_neon_vst1,
401                                                 Intrinsic::arm_neon_vst2,
402                                                 Intrinsic::arm_neon_vst3,
403                                                 Intrinsic::arm_neon_vst4};
404 
405       static const Intrinsic::ID StoreLaneInts[] = {
406         Intrinsic::arm_neon_vst2lane, Intrinsic::arm_neon_vst3lane,
407         Intrinsic::arm_neon_vst4lane
408       };
409 
410       auto fArgs = F->getFunctionType()->params();
411       Type *Tys[] = {fArgs[0], fArgs[1]};
412       if (Name.find("lane") == StringRef::npos)
413         NewFn = Intrinsic::getDeclaration(F->getParent(),
414                                           StoreInts[fArgs.size() - 3], Tys);
415       else
416         NewFn = Intrinsic::getDeclaration(F->getParent(),
417                                           StoreLaneInts[fArgs.size() - 5], Tys);
418       return true;
419     }
420     if (Name == "aarch64.thread.pointer" || Name == "arm.thread.pointer") {
421       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::thread_pointer);
422       return true;
423     }
424     break;
425   }
426 
427   case 'c': {
428     if (Name.startswith("ctlz.") && F->arg_size() == 1) {
429       rename(F);
430       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
431                                         F->arg_begin()->getType());
432       return true;
433     }
434     if (Name.startswith("cttz.") && F->arg_size() == 1) {
435       rename(F);
436       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::cttz,
437                                         F->arg_begin()->getType());
438       return true;
439     }
440     break;
441   }
442   case 'd': {
443     if (Name == "dbg.value" && F->arg_size() == 4) {
444       rename(F);
445       NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::dbg_value);
446       return true;
447     }
448     break;
449   }
450   case 'i':
451   case 'l': {
452     bool IsLifetimeStart = Name.startswith("lifetime.start");
453     if (IsLifetimeStart || Name.startswith("invariant.start")) {
454       Intrinsic::ID ID = IsLifetimeStart ?
455         Intrinsic::lifetime_start : Intrinsic::invariant_start;
456       auto Args = F->getFunctionType()->params();
457       Type* ObjectPtr[1] = {Args[1]};
458       if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
459         rename(F);
460         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
461         return true;
462       }
463     }
464 
465     bool IsLifetimeEnd = Name.startswith("lifetime.end");
466     if (IsLifetimeEnd || Name.startswith("invariant.end")) {
467       Intrinsic::ID ID = IsLifetimeEnd ?
468         Intrinsic::lifetime_end : Intrinsic::invariant_end;
469 
470       auto Args = F->getFunctionType()->params();
471       Type* ObjectPtr[1] = {Args[IsLifetimeEnd ? 1 : 2]};
472       if (F->getName() != Intrinsic::getName(ID, ObjectPtr)) {
473         rename(F);
474         NewFn = Intrinsic::getDeclaration(F->getParent(), ID, ObjectPtr);
475         return true;
476       }
477     }
478     break;
479   }
480   case 'm': {
481     if (Name.startswith("masked.load.")) {
482       Type *Tys[] = { F->getReturnType(), F->arg_begin()->getType() };
483       if (F->getName() != Intrinsic::getName(Intrinsic::masked_load, Tys)) {
484         rename(F);
485         NewFn = Intrinsic::getDeclaration(F->getParent(),
486                                           Intrinsic::masked_load,
487                                           Tys);
488         return true;
489       }
490     }
491     if (Name.startswith("masked.store.")) {
492       auto Args = F->getFunctionType()->params();
493       Type *Tys[] = { Args[0], Args[1] };
494       if (F->getName() != Intrinsic::getName(Intrinsic::masked_store, Tys)) {
495         rename(F);
496         NewFn = Intrinsic::getDeclaration(F->getParent(),
497                                           Intrinsic::masked_store,
498                                           Tys);
499         return true;
500       }
501     }
502     // Renaming gather/scatter intrinsics with no address space overloading
503     // to the new overload which includes an address space
504     if (Name.startswith("masked.gather.")) {
505       Type *Tys[] = {F->getReturnType(), F->arg_begin()->getType()};
506       if (F->getName() != Intrinsic::getName(Intrinsic::masked_gather, Tys)) {
507         rename(F);
508         NewFn = Intrinsic::getDeclaration(F->getParent(),
509                                           Intrinsic::masked_gather, Tys);
510         return true;
511       }
512     }
513     if (Name.startswith("masked.scatter.")) {
514       auto Args = F->getFunctionType()->params();
515       Type *Tys[] = {Args[0], Args[1]};
516       if (F->getName() != Intrinsic::getName(Intrinsic::masked_scatter, Tys)) {
517         rename(F);
518         NewFn = Intrinsic::getDeclaration(F->getParent(),
519                                           Intrinsic::masked_scatter, Tys);
520         return true;
521       }
522     }
523     break;
524   }
525   case 'n': {
526     if (Name.startswith("nvvm.")) {
527       Name = Name.substr(5);
528 
529       // The following nvvm intrinsics correspond exactly to an LLVM intrinsic.
530       Intrinsic::ID IID = StringSwitch<Intrinsic::ID>(Name)
531                               .Cases("brev32", "brev64", Intrinsic::bitreverse)
532                               .Case("clz.i", Intrinsic::ctlz)
533                               .Case("popc.i", Intrinsic::ctpop)
534                               .Default(Intrinsic::not_intrinsic);
535       if (IID != Intrinsic::not_intrinsic && F->arg_size() == 1) {
536         NewFn = Intrinsic::getDeclaration(F->getParent(), IID,
537                                           {F->getReturnType()});
538         return true;
539       }
540 
541       // The following nvvm intrinsics correspond exactly to an LLVM idiom, but
542       // not to an intrinsic alone.  We expand them in UpgradeIntrinsicCall.
543       //
544       // TODO: We could add lohi.i2d.
545       bool Expand = StringSwitch<bool>(Name)
546                         .Cases("abs.i", "abs.ll", true)
547                         .Cases("clz.ll", "popc.ll", "h2f", true)
548                         .Cases("max.i", "max.ll", "max.ui", "max.ull", true)
549                         .Cases("min.i", "min.ll", "min.ui", "min.ull", true)
550                         .Default(false);
551       if (Expand) {
552         NewFn = nullptr;
553         return true;
554       }
555     }
556     break;
557   }
558   case 'o':
559     // We only need to change the name to match the mangling including the
560     // address space.
561     if (Name.startswith("objectsize.")) {
562       Type *Tys[2] = { F->getReturnType(), F->arg_begin()->getType() };
563       if (F->arg_size() == 2 ||
564           F->getName() != Intrinsic::getName(Intrinsic::objectsize, Tys)) {
565         rename(F);
566         NewFn = Intrinsic::getDeclaration(F->getParent(), Intrinsic::objectsize,
567                                           Tys);
568         return true;
569       }
570     }
571     break;
572 
573   case 's':
574     if (Name == "stackprotectorcheck") {
575       NewFn = nullptr;
576       return true;
577     }
578     break;
579 
580   case 'x':
581     if (UpgradeX86IntrinsicFunction(F, Name, NewFn))
582       return true;
583   }
584   // Remangle our intrinsic since we upgrade the mangling
585   auto Result = llvm::Intrinsic::remangleIntrinsicFunction(F);
586   if (Result != None) {
587     NewFn = Result.getValue();
588     return true;
589   }
590 
591   //  This may not belong here. This function is effectively being overloaded
592   //  to both detect an intrinsic which needs upgrading, and to provide the
593   //  upgraded form of the intrinsic. We should perhaps have two separate
594   //  functions for this.
595   return false;
596 }
597 
598 bool llvm::UpgradeIntrinsicFunction(Function *F, Function *&NewFn) {
599   NewFn = nullptr;
600   bool Upgraded = UpgradeIntrinsicFunction1(F, NewFn);
601   assert(F != NewFn && "Intrinsic function upgraded to the same function");
602 
603   // Upgrade intrinsic attributes.  This does not change the function.
604   if (NewFn)
605     F = NewFn;
606   if (Intrinsic::ID id = F->getIntrinsicID())
607     F->setAttributes(Intrinsic::getAttributes(F->getContext(), id));
608   return Upgraded;
609 }
610 
611 bool llvm::UpgradeGlobalVariable(GlobalVariable *GV) {
612   // Nothing to do yet.
613   return false;
614 }
615 
616 // Handles upgrading SSE2/AVX2/AVX512BW PSLLDQ intrinsics by converting them
617 // to byte shuffles.
618 static Value *UpgradeX86PSLLDQIntrinsics(IRBuilder<> &Builder,
619                                          Value *Op, unsigned Shift) {
620   Type *ResultTy = Op->getType();
621   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
622 
623   // Bitcast from a 64-bit element type to a byte element type.
624   Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
625   Op = Builder.CreateBitCast(Op, VecTy, "cast");
626 
627   // We'll be shuffling in zeroes.
628   Value *Res = Constant::getNullValue(VecTy);
629 
630   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
631   // we'll just return the zero vector.
632   if (Shift < 16) {
633     uint32_t Idxs[64];
634     // 256/512-bit version is split into 2/4 16-byte lanes.
635     for (unsigned l = 0; l != NumElts; l += 16)
636       for (unsigned i = 0; i != 16; ++i) {
637         unsigned Idx = NumElts + i - Shift;
638         if (Idx < NumElts)
639           Idx -= NumElts - 16; // end of lane, switch operand.
640         Idxs[l + i] = Idx + l;
641       }
642 
643     Res = Builder.CreateShuffleVector(Res, Op, makeArrayRef(Idxs, NumElts));
644   }
645 
646   // Bitcast back to a 64-bit element type.
647   return Builder.CreateBitCast(Res, ResultTy, "cast");
648 }
649 
650 // Handles upgrading SSE2/AVX2/AVX512BW PSRLDQ intrinsics by converting them
651 // to byte shuffles.
652 static Value *UpgradeX86PSRLDQIntrinsics(IRBuilder<> &Builder, Value *Op,
653                                          unsigned Shift) {
654   Type *ResultTy = Op->getType();
655   unsigned NumElts = ResultTy->getVectorNumElements() * 8;
656 
657   // Bitcast from a 64-bit element type to a byte element type.
658   Type *VecTy = VectorType::get(Builder.getInt8Ty(), NumElts);
659   Op = Builder.CreateBitCast(Op, VecTy, "cast");
660 
661   // We'll be shuffling in zeroes.
662   Value *Res = Constant::getNullValue(VecTy);
663 
664   // If shift is less than 16, emit a shuffle to move the bytes. Otherwise,
665   // we'll just return the zero vector.
666   if (Shift < 16) {
667     uint32_t Idxs[64];
668     // 256/512-bit version is split into 2/4 16-byte lanes.
669     for (unsigned l = 0; l != NumElts; l += 16)
670       for (unsigned i = 0; i != 16; ++i) {
671         unsigned Idx = i + Shift;
672         if (Idx >= 16)
673           Idx += NumElts - 16; // end of lane, switch operand.
674         Idxs[l + i] = Idx + l;
675       }
676 
677     Res = Builder.CreateShuffleVector(Op, Res, makeArrayRef(Idxs, NumElts));
678   }
679 
680   // Bitcast back to a 64-bit element type.
681   return Builder.CreateBitCast(Res, ResultTy, "cast");
682 }
683 
684 static Value *getX86MaskVec(IRBuilder<> &Builder, Value *Mask,
685                             unsigned NumElts) {
686   llvm::VectorType *MaskTy = llvm::VectorType::get(Builder.getInt1Ty(),
687                              cast<IntegerType>(Mask->getType())->getBitWidth());
688   Mask = Builder.CreateBitCast(Mask, MaskTy);
689 
690   // If we have less than 8 elements, then the starting mask was an i8 and
691   // we need to extract down to the right number of elements.
692   if (NumElts < 8) {
693     uint32_t Indices[4];
694     for (unsigned i = 0; i != NumElts; ++i)
695       Indices[i] = i;
696     Mask = Builder.CreateShuffleVector(Mask, Mask,
697                                        makeArrayRef(Indices, NumElts),
698                                        "extract");
699   }
700 
701   return Mask;
702 }
703 
704 static Value *EmitX86Select(IRBuilder<> &Builder, Value *Mask,
705                             Value *Op0, Value *Op1) {
706   // If the mask is all ones just emit the align operation.
707   if (const auto *C = dyn_cast<Constant>(Mask))
708     if (C->isAllOnesValue())
709       return Op0;
710 
711   Mask = getX86MaskVec(Builder, Mask, Op0->getType()->getVectorNumElements());
712   return Builder.CreateSelect(Mask, Op0, Op1);
713 }
714 
715 // Handle autoupgrade for masked PALIGNR and VALIGND/Q intrinsics.
716 // PALIGNR handles large immediates by shifting while VALIGN masks the immediate
717 // so we need to handle both cases. VALIGN also doesn't have 128-bit lanes.
718 static Value *UpgradeX86ALIGNIntrinsics(IRBuilder<> &Builder, Value *Op0,
719                                         Value *Op1, Value *Shift,
720                                         Value *Passthru, Value *Mask,
721                                         bool IsVALIGN) {
722   unsigned ShiftVal = cast<llvm::ConstantInt>(Shift)->getZExtValue();
723 
724   unsigned NumElts = Op0->getType()->getVectorNumElements();
725   assert((IsVALIGN || NumElts % 16 == 0) && "Illegal NumElts for PALIGNR!");
726   assert((!IsVALIGN || NumElts <= 16) && "NumElts too large for VALIGN!");
727   assert(isPowerOf2_32(NumElts) && "NumElts not a power of 2!");
728 
729   // Mask the immediate for VALIGN.
730   if (IsVALIGN)
731     ShiftVal &= (NumElts - 1);
732 
733   // If palignr is shifting the pair of vectors more than the size of two
734   // lanes, emit zero.
735   if (ShiftVal >= 32)
736     return llvm::Constant::getNullValue(Op0->getType());
737 
738   // If palignr is shifting the pair of input vectors more than one lane,
739   // but less than two lanes, convert to shifting in zeroes.
740   if (ShiftVal > 16) {
741     ShiftVal -= 16;
742     Op1 = Op0;
743     Op0 = llvm::Constant::getNullValue(Op0->getType());
744   }
745 
746   uint32_t Indices[64];
747   // 256-bit palignr operates on 128-bit lanes so we need to handle that
748   for (unsigned l = 0; l < NumElts; l += 16) {
749     for (unsigned i = 0; i != 16; ++i) {
750       unsigned Idx = ShiftVal + i;
751       if (!IsVALIGN && Idx >= 16) // Disable wrap for VALIGN.
752         Idx += NumElts - 16; // End of lane, switch operand.
753       Indices[l + i] = Idx + l;
754     }
755   }
756 
757   Value *Align = Builder.CreateShuffleVector(Op1, Op0,
758                                              makeArrayRef(Indices, NumElts),
759                                              "palignr");
760 
761   return EmitX86Select(Builder, Mask, Align, Passthru);
762 }
763 
764 static Value *UpgradeMaskedStore(IRBuilder<> &Builder,
765                                  Value *Ptr, Value *Data, Value *Mask,
766                                  bool Aligned) {
767   // Cast the pointer to the right type.
768   Ptr = Builder.CreateBitCast(Ptr,
769                               llvm::PointerType::getUnqual(Data->getType()));
770   unsigned Align =
771     Aligned ? cast<VectorType>(Data->getType())->getBitWidth() / 8 : 1;
772 
773   // If the mask is all ones just emit a regular store.
774   if (const auto *C = dyn_cast<Constant>(Mask))
775     if (C->isAllOnesValue())
776       return Builder.CreateAlignedStore(Data, Ptr, Align);
777 
778   // Convert the mask from an integer type to a vector of i1.
779   unsigned NumElts = Data->getType()->getVectorNumElements();
780   Mask = getX86MaskVec(Builder, Mask, NumElts);
781   return Builder.CreateMaskedStore(Data, Ptr, Align, Mask);
782 }
783 
784 static Value *UpgradeMaskedLoad(IRBuilder<> &Builder,
785                                 Value *Ptr, Value *Passthru, Value *Mask,
786                                 bool Aligned) {
787   // Cast the pointer to the right type.
788   Ptr = Builder.CreateBitCast(Ptr,
789                              llvm::PointerType::getUnqual(Passthru->getType()));
790   unsigned Align =
791     Aligned ? cast<VectorType>(Passthru->getType())->getBitWidth() / 8 : 1;
792 
793   // If the mask is all ones just emit a regular store.
794   if (const auto *C = dyn_cast<Constant>(Mask))
795     if (C->isAllOnesValue())
796       return Builder.CreateAlignedLoad(Ptr, Align);
797 
798   // Convert the mask from an integer type to a vector of i1.
799   unsigned NumElts = Passthru->getType()->getVectorNumElements();
800   Mask = getX86MaskVec(Builder, Mask, NumElts);
801   return Builder.CreateMaskedLoad(Ptr, Align, Mask, Passthru);
802 }
803 
804 static Value *upgradeAbs(IRBuilder<> &Builder, CallInst &CI) {
805   Value *Op0 = CI.getArgOperand(0);
806   llvm::Type *Ty = Op0->getType();
807   Value *Zero = llvm::Constant::getNullValue(Ty);
808   Value *Cmp = Builder.CreateICmp(ICmpInst::ICMP_SGT, Op0, Zero);
809   Value *Neg = Builder.CreateNeg(Op0);
810   Value *Res = Builder.CreateSelect(Cmp, Op0, Neg);
811 
812   if (CI.getNumArgOperands() == 3)
813     Res = EmitX86Select(Builder,CI.getArgOperand(2), Res, CI.getArgOperand(1));
814 
815   return Res;
816 }
817 
818 static Value *upgradeIntMinMax(IRBuilder<> &Builder, CallInst &CI,
819                                ICmpInst::Predicate Pred) {
820   Value *Op0 = CI.getArgOperand(0);
821   Value *Op1 = CI.getArgOperand(1);
822   Value *Cmp = Builder.CreateICmp(Pred, Op0, Op1);
823   Value *Res = Builder.CreateSelect(Cmp, Op0, Op1);
824 
825   if (CI.getNumArgOperands() == 4)
826     Res = EmitX86Select(Builder, CI.getArgOperand(3), Res, CI.getArgOperand(2));
827 
828   return Res;
829 }
830 
831 static Value *upgradeMaskedCompare(IRBuilder<> &Builder, CallInst &CI,
832                                    unsigned CC, bool Signed) {
833   Value *Op0 = CI.getArgOperand(0);
834   unsigned NumElts = Op0->getType()->getVectorNumElements();
835 
836   Value *Cmp;
837   if (CC == 3) {
838     Cmp = Constant::getNullValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
839   } else if (CC == 7) {
840     Cmp = Constant::getAllOnesValue(llvm::VectorType::get(Builder.getInt1Ty(), NumElts));
841   } else {
842     ICmpInst::Predicate Pred;
843     switch (CC) {
844     default: llvm_unreachable("Unknown condition code");
845     case 0: Pred = ICmpInst::ICMP_EQ;  break;
846     case 1: Pred = Signed ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT; break;
847     case 2: Pred = Signed ? ICmpInst::ICMP_SLE : ICmpInst::ICMP_ULE; break;
848     case 4: Pred = ICmpInst::ICMP_NE;  break;
849     case 5: Pred = Signed ? ICmpInst::ICMP_SGE : ICmpInst::ICMP_UGE; break;
850     case 6: Pred = Signed ? ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT; break;
851     }
852     Cmp = Builder.CreateICmp(Pred, Op0, CI.getArgOperand(1));
853   }
854 
855   Value *Mask = CI.getArgOperand(CI.getNumArgOperands() - 1);
856   const auto *C = dyn_cast<Constant>(Mask);
857   if (!C || !C->isAllOnesValue())
858     Cmp = Builder.CreateAnd(Cmp, getX86MaskVec(Builder, Mask, NumElts));
859 
860   if (NumElts < 8) {
861     uint32_t Indices[8];
862     for (unsigned i = 0; i != NumElts; ++i)
863       Indices[i] = i;
864     for (unsigned i = NumElts; i != 8; ++i)
865       Indices[i] = NumElts + i % NumElts;
866     Cmp = Builder.CreateShuffleVector(Cmp,
867                                       Constant::getNullValue(Cmp->getType()),
868                                       Indices);
869   }
870   return Builder.CreateBitCast(Cmp, IntegerType::get(CI.getContext(),
871                                                      std::max(NumElts, 8U)));
872 }
873 
874 // Replace a masked intrinsic with an older unmasked intrinsic.
875 static Value *UpgradeX86MaskedShift(IRBuilder<> &Builder, CallInst &CI,
876                                     Intrinsic::ID IID) {
877   Function *F = CI.getCalledFunction();
878   Function *Intrin = Intrinsic::getDeclaration(F->getParent(), IID);
879   Value *Rep = Builder.CreateCall(Intrin,
880                                  { CI.getArgOperand(0), CI.getArgOperand(1) });
881   return EmitX86Select(Builder, CI.getArgOperand(3), Rep, CI.getArgOperand(2));
882 }
883 
884 static Value* upgradeMaskedMove(IRBuilder<> &Builder, CallInst &CI) {
885   Value* A = CI.getArgOperand(0);
886   Value* B = CI.getArgOperand(1);
887   Value* Src = CI.getArgOperand(2);
888   Value* Mask = CI.getArgOperand(3);
889 
890   Value* AndNode = Builder.CreateAnd(Mask, APInt(8, 1));
891   Value* Cmp = Builder.CreateIsNotNull(AndNode);
892   Value* Extract1 = Builder.CreateExtractElement(B, (uint64_t)0);
893   Value* Extract2 = Builder.CreateExtractElement(Src, (uint64_t)0);
894   Value* Select = Builder.CreateSelect(Cmp, Extract1, Extract2);
895   return Builder.CreateInsertElement(A, Select, (uint64_t)0);
896 }
897 
898 
899 static Value* UpgradeMaskToInt(IRBuilder<> &Builder, CallInst &CI) {
900   Value* Op = CI.getArgOperand(0);
901   Type* ReturnOp = CI.getType();
902   unsigned NumElts = CI.getType()->getVectorNumElements();
903   Value *Mask = getX86MaskVec(Builder, Op, NumElts);
904   return Builder.CreateSExt(Mask, ReturnOp, "vpmovm2");
905 }
906 
907 /// Upgrade a call to an old intrinsic. All argument and return casting must be
908 /// provided to seamlessly integrate with existing context.
909 void llvm::UpgradeIntrinsicCall(CallInst *CI, Function *NewFn) {
910   Function *F = CI->getCalledFunction();
911   LLVMContext &C = CI->getContext();
912   IRBuilder<> Builder(C);
913   Builder.SetInsertPoint(CI->getParent(), CI->getIterator());
914 
915   assert(F && "Intrinsic call is not direct?");
916 
917   if (!NewFn) {
918     // Get the Function's name.
919     StringRef Name = F->getName();
920 
921     assert(Name.startswith("llvm.") && "Intrinsic doesn't start with 'llvm.'");
922     Name = Name.substr(5);
923 
924     bool IsX86 = Name.startswith("x86.");
925     if (IsX86)
926       Name = Name.substr(4);
927     bool IsNVVM = Name.startswith("nvvm.");
928     if (IsNVVM)
929       Name = Name.substr(5);
930 
931     if (IsX86 && Name.startswith("sse4a.movnt.")) {
932       Module *M = F->getParent();
933       SmallVector<Metadata *, 1> Elts;
934       Elts.push_back(
935           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
936       MDNode *Node = MDNode::get(C, Elts);
937 
938       Value *Arg0 = CI->getArgOperand(0);
939       Value *Arg1 = CI->getArgOperand(1);
940 
941       // Nontemporal (unaligned) store of the 0'th element of the float/double
942       // vector.
943       Type *SrcEltTy = cast<VectorType>(Arg1->getType())->getElementType();
944       PointerType *EltPtrTy = PointerType::getUnqual(SrcEltTy);
945       Value *Addr = Builder.CreateBitCast(Arg0, EltPtrTy, "cast");
946       Value *Extract =
947           Builder.CreateExtractElement(Arg1, (uint64_t)0, "extractelement");
948 
949       StoreInst *SI = Builder.CreateAlignedStore(Extract, Addr, 1);
950       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
951 
952       // Remove intrinsic.
953       CI->eraseFromParent();
954       return;
955     }
956 
957     if (IsX86 && (Name.startswith("avx.movnt.") ||
958                   Name.startswith("avx512.storent."))) {
959       Module *M = F->getParent();
960       SmallVector<Metadata *, 1> Elts;
961       Elts.push_back(
962           ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
963       MDNode *Node = MDNode::get(C, Elts);
964 
965       Value *Arg0 = CI->getArgOperand(0);
966       Value *Arg1 = CI->getArgOperand(1);
967 
968       // Convert the type of the pointer to a pointer to the stored type.
969       Value *BC = Builder.CreateBitCast(Arg0,
970                                         PointerType::getUnqual(Arg1->getType()),
971                                         "cast");
972       VectorType *VTy = cast<VectorType>(Arg1->getType());
973       StoreInst *SI = Builder.CreateAlignedStore(Arg1, BC,
974                                                  VTy->getBitWidth() / 8);
975       SI->setMetadata(M->getMDKindID("nontemporal"), Node);
976 
977       // Remove intrinsic.
978       CI->eraseFromParent();
979       return;
980     }
981 
982     if (IsX86 && Name == "sse2.storel.dq") {
983       Value *Arg0 = CI->getArgOperand(0);
984       Value *Arg1 = CI->getArgOperand(1);
985 
986       Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
987       Value *BC0 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
988       Value *Elt = Builder.CreateExtractElement(BC0, (uint64_t)0);
989       Value *BC = Builder.CreateBitCast(Arg0,
990                                         PointerType::getUnqual(Elt->getType()),
991                                         "cast");
992       Builder.CreateAlignedStore(Elt, BC, 1);
993 
994       // Remove intrinsic.
995       CI->eraseFromParent();
996       return;
997     }
998 
999     if (IsX86 && (Name.startswith("sse.storeu.") ||
1000                   Name.startswith("sse2.storeu.") ||
1001                   Name.startswith("avx.storeu."))) {
1002       Value *Arg0 = CI->getArgOperand(0);
1003       Value *Arg1 = CI->getArgOperand(1);
1004 
1005       Arg0 = Builder.CreateBitCast(Arg0,
1006                                    PointerType::getUnqual(Arg1->getType()),
1007                                    "cast");
1008       Builder.CreateAlignedStore(Arg1, Arg0, 1);
1009 
1010       // Remove intrinsic.
1011       CI->eraseFromParent();
1012       return;
1013     }
1014 
1015     if (IsX86 && (Name.startswith("avx512.mask.store"))) {
1016       // "avx512.mask.storeu." or "avx512.mask.store."
1017       bool Aligned = Name[17] != 'u'; // "avx512.mask.storeu".
1018       UpgradeMaskedStore(Builder, CI->getArgOperand(0), CI->getArgOperand(1),
1019                          CI->getArgOperand(2), Aligned);
1020 
1021       // Remove intrinsic.
1022       CI->eraseFromParent();
1023       return;
1024     }
1025 
1026     Value *Rep;
1027     // Upgrade packed integer vector compare intrinsics to compare instructions.
1028     if (IsX86 && (Name.startswith("sse2.pcmp") ||
1029                   Name.startswith("avx2.pcmp"))) {
1030       // "sse2.pcpmpeq." "sse2.pcmpgt." "avx2.pcmpeq." or "avx2.pcmpgt."
1031       bool CmpEq = Name[9] == 'e';
1032       Rep = Builder.CreateICmp(CmpEq ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_SGT,
1033                                CI->getArgOperand(0), CI->getArgOperand(1));
1034       Rep = Builder.CreateSExt(Rep, CI->getType(), "");
1035     } else if (IsX86 && (Name.startswith("avx512.mask.pbroadcast"))){
1036       unsigned NumElts =
1037           CI->getArgOperand(1)->getType()->getVectorNumElements();
1038       Rep = Builder.CreateVectorSplat(NumElts, CI->getArgOperand(0));
1039       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1040                           CI->getArgOperand(1));
1041     } else if (IsX86 && (Name == "sse.add.ss" || Name == "sse2.add.sd")) {
1042       Type *I32Ty = Type::getInt32Ty(C);
1043       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1044                                                  ConstantInt::get(I32Ty, 0));
1045       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1046                                                  ConstantInt::get(I32Ty, 0));
1047       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1048                                         Builder.CreateFAdd(Elt0, Elt1),
1049                                         ConstantInt::get(I32Ty, 0));
1050     } else if (IsX86 && (Name == "sse.sub.ss" || Name == "sse2.sub.sd")) {
1051       Type *I32Ty = Type::getInt32Ty(C);
1052       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1053                                                  ConstantInt::get(I32Ty, 0));
1054       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1055                                                  ConstantInt::get(I32Ty, 0));
1056       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1057                                         Builder.CreateFSub(Elt0, Elt1),
1058                                         ConstantInt::get(I32Ty, 0));
1059     } else if (IsX86 && (Name == "sse.mul.ss" || Name == "sse2.mul.sd")) {
1060       Type *I32Ty = Type::getInt32Ty(C);
1061       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1062                                                  ConstantInt::get(I32Ty, 0));
1063       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1064                                                  ConstantInt::get(I32Ty, 0));
1065       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1066                                         Builder.CreateFMul(Elt0, Elt1),
1067                                         ConstantInt::get(I32Ty, 0));
1068     } else if (IsX86 && (Name == "sse.div.ss" || Name == "sse2.div.sd")) {
1069       Type *I32Ty = Type::getInt32Ty(C);
1070       Value *Elt0 = Builder.CreateExtractElement(CI->getArgOperand(0),
1071                                                  ConstantInt::get(I32Ty, 0));
1072       Value *Elt1 = Builder.CreateExtractElement(CI->getArgOperand(1),
1073                                                  ConstantInt::get(I32Ty, 0));
1074       Rep = Builder.CreateInsertElement(CI->getArgOperand(0),
1075                                         Builder.CreateFDiv(Elt0, Elt1),
1076                                         ConstantInt::get(I32Ty, 0));
1077     } else if (IsX86 && Name.startswith("avx512.mask.pcmp")) {
1078       // "avx512.mask.pcmpeq." or "avx512.mask.pcmpgt."
1079       bool CmpEq = Name[16] == 'e';
1080       Rep = upgradeMaskedCompare(Builder, *CI, CmpEq ? 0 : 6, true);
1081     } else if (IsX86 && Name.startswith("avx512.mask.cmp")) {
1082       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1083       Rep = upgradeMaskedCompare(Builder, *CI, Imm, true);
1084     } else if (IsX86 && Name.startswith("avx512.mask.ucmp")) {
1085       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1086       Rep = upgradeMaskedCompare(Builder, *CI, Imm, false);
1087     } else if(IsX86 && (Name == "ssse3.pabs.b.128" ||
1088                         Name == "ssse3.pabs.w.128" ||
1089                         Name == "ssse3.pabs.d.128" ||
1090                         Name.startswith("avx2.pabs") ||
1091                         Name.startswith("avx512.mask.pabs"))) {
1092       Rep = upgradeAbs(Builder, *CI);
1093     } else if (IsX86 && (Name == "sse41.pmaxsb" ||
1094                          Name == "sse2.pmaxs.w" ||
1095                          Name == "sse41.pmaxsd" ||
1096                          Name.startswith("avx2.pmaxs") ||
1097                          Name.startswith("avx512.mask.pmaxs"))) {
1098       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SGT);
1099     } else if (IsX86 && (Name == "sse2.pmaxu.b" ||
1100                          Name == "sse41.pmaxuw" ||
1101                          Name == "sse41.pmaxud" ||
1102                          Name.startswith("avx2.pmaxu") ||
1103                          Name.startswith("avx512.mask.pmaxu"))) {
1104       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_UGT);
1105     } else if (IsX86 && (Name == "sse41.pminsb" ||
1106                          Name == "sse2.pmins.w" ||
1107                          Name == "sse41.pminsd" ||
1108                          Name.startswith("avx2.pmins") ||
1109                          Name.startswith("avx512.mask.pmins"))) {
1110       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_SLT);
1111     } else if (IsX86 && (Name == "sse2.pminu.b" ||
1112                          Name == "sse41.pminuw" ||
1113                          Name == "sse41.pminud" ||
1114                          Name.startswith("avx2.pminu") ||
1115                          Name.startswith("avx512.mask.pminu"))) {
1116       Rep = upgradeIntMinMax(Builder, *CI, ICmpInst::ICMP_ULT);
1117     } else if (IsX86 && (Name == "sse2.cvtdq2pd" ||
1118                          Name == "sse2.cvtps2pd" ||
1119                          Name == "avx.cvtdq2.pd.256" ||
1120                          Name == "avx.cvt.ps2.pd.256" ||
1121                          Name.startswith("avx512.mask.cvtdq2pd.") ||
1122                          Name.startswith("avx512.mask.cvtudq2pd."))) {
1123       // Lossless i32/float to double conversion.
1124       // Extract the bottom elements if necessary and convert to double vector.
1125       Value *Src = CI->getArgOperand(0);
1126       VectorType *SrcTy = cast<VectorType>(Src->getType());
1127       VectorType *DstTy = cast<VectorType>(CI->getType());
1128       Rep = CI->getArgOperand(0);
1129 
1130       unsigned NumDstElts = DstTy->getNumElements();
1131       if (NumDstElts < SrcTy->getNumElements()) {
1132         assert(NumDstElts == 2 && "Unexpected vector size");
1133         uint32_t ShuffleMask[2] = { 0, 1 };
1134         Rep = Builder.CreateShuffleVector(Rep, UndefValue::get(SrcTy),
1135                                           ShuffleMask);
1136       }
1137 
1138       bool SInt2Double = (StringRef::npos != Name.find("cvtdq2"));
1139       bool UInt2Double = (StringRef::npos != Name.find("cvtudq2"));
1140       if (SInt2Double)
1141         Rep = Builder.CreateSIToFP(Rep, DstTy, "cvtdq2pd");
1142       else if (UInt2Double)
1143         Rep = Builder.CreateUIToFP(Rep, DstTy, "cvtudq2pd");
1144       else
1145         Rep = Builder.CreateFPExt(Rep, DstTy, "cvtps2pd");
1146 
1147       if (CI->getNumArgOperands() == 3)
1148         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1149                             CI->getArgOperand(1));
1150     } else if (IsX86 && (Name.startswith("avx512.mask.loadu."))) {
1151       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
1152                               CI->getArgOperand(1), CI->getArgOperand(2),
1153                               /*Aligned*/false);
1154     } else if (IsX86 && (Name.startswith("avx512.mask.load."))) {
1155       Rep = UpgradeMaskedLoad(Builder, CI->getArgOperand(0),
1156                               CI->getArgOperand(1),CI->getArgOperand(2),
1157                               /*Aligned*/true);
1158     } else if (IsX86 && Name.startswith("xop.vpcom")) {
1159       Intrinsic::ID intID;
1160       if (Name.endswith("ub"))
1161         intID = Intrinsic::x86_xop_vpcomub;
1162       else if (Name.endswith("uw"))
1163         intID = Intrinsic::x86_xop_vpcomuw;
1164       else if (Name.endswith("ud"))
1165         intID = Intrinsic::x86_xop_vpcomud;
1166       else if (Name.endswith("uq"))
1167         intID = Intrinsic::x86_xop_vpcomuq;
1168       else if (Name.endswith("b"))
1169         intID = Intrinsic::x86_xop_vpcomb;
1170       else if (Name.endswith("w"))
1171         intID = Intrinsic::x86_xop_vpcomw;
1172       else if (Name.endswith("d"))
1173         intID = Intrinsic::x86_xop_vpcomd;
1174       else if (Name.endswith("q"))
1175         intID = Intrinsic::x86_xop_vpcomq;
1176       else
1177         llvm_unreachable("Unknown suffix");
1178 
1179       Name = Name.substr(9); // strip off "xop.vpcom"
1180       unsigned Imm;
1181       if (Name.startswith("lt"))
1182         Imm = 0;
1183       else if (Name.startswith("le"))
1184         Imm = 1;
1185       else if (Name.startswith("gt"))
1186         Imm = 2;
1187       else if (Name.startswith("ge"))
1188         Imm = 3;
1189       else if (Name.startswith("eq"))
1190         Imm = 4;
1191       else if (Name.startswith("ne"))
1192         Imm = 5;
1193       else if (Name.startswith("false"))
1194         Imm = 6;
1195       else if (Name.startswith("true"))
1196         Imm = 7;
1197       else
1198         llvm_unreachable("Unknown condition");
1199 
1200       Function *VPCOM = Intrinsic::getDeclaration(F->getParent(), intID);
1201       Rep =
1202           Builder.CreateCall(VPCOM, {CI->getArgOperand(0), CI->getArgOperand(1),
1203                                      Builder.getInt8(Imm)});
1204     } else if (IsX86 && Name.startswith("xop.vpcmov")) {
1205       Value *Sel = CI->getArgOperand(2);
1206       Value *NotSel = Builder.CreateNot(Sel);
1207       Value *Sel0 = Builder.CreateAnd(CI->getArgOperand(0), Sel);
1208       Value *Sel1 = Builder.CreateAnd(CI->getArgOperand(1), NotSel);
1209       Rep = Builder.CreateOr(Sel0, Sel1);
1210     } else if (IsX86 && Name == "sse42.crc32.64.8") {
1211       Function *CRC32 = Intrinsic::getDeclaration(F->getParent(),
1212                                                Intrinsic::x86_sse42_crc32_32_8);
1213       Value *Trunc0 = Builder.CreateTrunc(CI->getArgOperand(0), Type::getInt32Ty(C));
1214       Rep = Builder.CreateCall(CRC32, {Trunc0, CI->getArgOperand(1)});
1215       Rep = Builder.CreateZExt(Rep, CI->getType(), "");
1216     } else if (IsX86 && Name.startswith("avx.vbroadcast.s")) {
1217       // Replace broadcasts with a series of insertelements.
1218       Type *VecTy = CI->getType();
1219       Type *EltTy = VecTy->getVectorElementType();
1220       unsigned EltNum = VecTy->getVectorNumElements();
1221       Value *Cast = Builder.CreateBitCast(CI->getArgOperand(0),
1222                                           EltTy->getPointerTo());
1223       Value *Load = Builder.CreateLoad(EltTy, Cast);
1224       Type *I32Ty = Type::getInt32Ty(C);
1225       Rep = UndefValue::get(VecTy);
1226       for (unsigned I = 0; I < EltNum; ++I)
1227         Rep = Builder.CreateInsertElement(Rep, Load,
1228                                           ConstantInt::get(I32Ty, I));
1229     } else if (IsX86 && (Name.startswith("sse41.pmovsx") ||
1230                          Name.startswith("sse41.pmovzx") ||
1231                          Name.startswith("avx2.pmovsx") ||
1232                          Name.startswith("avx2.pmovzx") ||
1233                          Name.startswith("avx512.mask.pmovsx") ||
1234                          Name.startswith("avx512.mask.pmovzx"))) {
1235       VectorType *SrcTy = cast<VectorType>(CI->getArgOperand(0)->getType());
1236       VectorType *DstTy = cast<VectorType>(CI->getType());
1237       unsigned NumDstElts = DstTy->getNumElements();
1238 
1239       // Extract a subvector of the first NumDstElts lanes and sign/zero extend.
1240       SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
1241       for (unsigned i = 0; i != NumDstElts; ++i)
1242         ShuffleMask[i] = i;
1243 
1244       Value *SV = Builder.CreateShuffleVector(
1245           CI->getArgOperand(0), UndefValue::get(SrcTy), ShuffleMask);
1246 
1247       bool DoSext = (StringRef::npos != Name.find("pmovsx"));
1248       Rep = DoSext ? Builder.CreateSExt(SV, DstTy)
1249                    : Builder.CreateZExt(SV, DstTy);
1250       // If there are 3 arguments, it's a masked intrinsic so we need a select.
1251       if (CI->getNumArgOperands() == 3)
1252         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1253                             CI->getArgOperand(1));
1254     } else if (IsX86 && (Name.startswith("avx.vbroadcastf128") ||
1255                          Name == "avx2.vbroadcasti128")) {
1256       // Replace vbroadcastf128/vbroadcasti128 with a vector load+shuffle.
1257       Type *EltTy = CI->getType()->getVectorElementType();
1258       unsigned NumSrcElts = 128 / EltTy->getPrimitiveSizeInBits();
1259       Type *VT = VectorType::get(EltTy, NumSrcElts);
1260       Value *Op = Builder.CreatePointerCast(CI->getArgOperand(0),
1261                                             PointerType::getUnqual(VT));
1262       Value *Load = Builder.CreateAlignedLoad(Op, 1);
1263       if (NumSrcElts == 2)
1264         Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
1265                                           { 0, 1, 0, 1 });
1266       else
1267         Rep = Builder.CreateShuffleVector(Load, UndefValue::get(Load->getType()),
1268                                           { 0, 1, 2, 3, 0, 1, 2, 3 });
1269     } else if (IsX86 && (Name.startswith("avx512.mask.broadcastf") ||
1270                          Name.startswith("avx512.mask.broadcasti"))) {
1271       unsigned NumSrcElts =
1272                         CI->getArgOperand(0)->getType()->getVectorNumElements();
1273       unsigned NumDstElts = CI->getType()->getVectorNumElements();
1274 
1275       SmallVector<uint32_t, 8> ShuffleMask(NumDstElts);
1276       for (unsigned i = 0; i != NumDstElts; ++i)
1277         ShuffleMask[i] = i % NumSrcElts;
1278 
1279       Rep = Builder.CreateShuffleVector(CI->getArgOperand(0),
1280                                         CI->getArgOperand(0),
1281                                         ShuffleMask);
1282       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1283                           CI->getArgOperand(1));
1284     } else if (IsX86 && (Name.startswith("avx2.pbroadcast") ||
1285                          Name.startswith("avx2.vbroadcast") ||
1286                          Name.startswith("avx512.pbroadcast") ||
1287                          Name.startswith("avx512.mask.broadcast.s"))) {
1288       // Replace vp?broadcasts with a vector shuffle.
1289       Value *Op = CI->getArgOperand(0);
1290       unsigned NumElts = CI->getType()->getVectorNumElements();
1291       Type *MaskTy = VectorType::get(Type::getInt32Ty(C), NumElts);
1292       Rep = Builder.CreateShuffleVector(Op, UndefValue::get(Op->getType()),
1293                                         Constant::getNullValue(MaskTy));
1294 
1295       if (CI->getNumArgOperands() == 3)
1296         Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1297                             CI->getArgOperand(1));
1298     } else if (IsX86 && Name.startswith("avx512.mask.palignr.")) {
1299       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
1300                                       CI->getArgOperand(1),
1301                                       CI->getArgOperand(2),
1302                                       CI->getArgOperand(3),
1303                                       CI->getArgOperand(4),
1304                                       false);
1305     } else if (IsX86 && Name.startswith("avx512.mask.valign.")) {
1306       Rep = UpgradeX86ALIGNIntrinsics(Builder, CI->getArgOperand(0),
1307                                       CI->getArgOperand(1),
1308                                       CI->getArgOperand(2),
1309                                       CI->getArgOperand(3),
1310                                       CI->getArgOperand(4),
1311                                       true);
1312     } else if (IsX86 && (Name == "sse2.psll.dq" ||
1313                          Name == "avx2.psll.dq")) {
1314       // 128/256-bit shift left specified in bits.
1315       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1316       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0),
1317                                        Shift / 8); // Shift is in bits.
1318     } else if (IsX86 && (Name == "sse2.psrl.dq" ||
1319                          Name == "avx2.psrl.dq")) {
1320       // 128/256-bit shift right specified in bits.
1321       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1322       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0),
1323                                        Shift / 8); // Shift is in bits.
1324     } else if (IsX86 && (Name == "sse2.psll.dq.bs" ||
1325                          Name == "avx2.psll.dq.bs" ||
1326                          Name == "avx512.psll.dq.512")) {
1327       // 128/256/512-bit shift left specified in bytes.
1328       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1329       Rep = UpgradeX86PSLLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
1330     } else if (IsX86 && (Name == "sse2.psrl.dq.bs" ||
1331                          Name == "avx2.psrl.dq.bs" ||
1332                          Name == "avx512.psrl.dq.512")) {
1333       // 128/256/512-bit shift right specified in bytes.
1334       unsigned Shift = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1335       Rep = UpgradeX86PSRLDQIntrinsics(Builder, CI->getArgOperand(0), Shift);
1336     } else if (IsX86 && (Name == "sse41.pblendw" ||
1337                          Name.startswith("sse41.blendp") ||
1338                          Name.startswith("avx.blend.p") ||
1339                          Name == "avx2.pblendw" ||
1340                          Name.startswith("avx2.pblendd."))) {
1341       Value *Op0 = CI->getArgOperand(0);
1342       Value *Op1 = CI->getArgOperand(1);
1343       unsigned Imm = cast <ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1344       VectorType *VecTy = cast<VectorType>(CI->getType());
1345       unsigned NumElts = VecTy->getNumElements();
1346 
1347       SmallVector<uint32_t, 16> Idxs(NumElts);
1348       for (unsigned i = 0; i != NumElts; ++i)
1349         Idxs[i] = ((Imm >> (i%8)) & 1) ? i + NumElts : i;
1350 
1351       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1352     } else if (IsX86 && (Name.startswith("avx.vinsertf128.") ||
1353                          Name == "avx2.vinserti128" ||
1354                          Name.startswith("avx512.mask.insert"))) {
1355       Value *Op0 = CI->getArgOperand(0);
1356       Value *Op1 = CI->getArgOperand(1);
1357       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1358       unsigned DstNumElts = CI->getType()->getVectorNumElements();
1359       unsigned SrcNumElts = Op1->getType()->getVectorNumElements();
1360       unsigned Scale = DstNumElts / SrcNumElts;
1361 
1362       // Mask off the high bits of the immediate value; hardware ignores those.
1363       Imm = Imm % Scale;
1364 
1365       // Extend the second operand into a vector the size of the destination.
1366       Value *UndefV = UndefValue::get(Op1->getType());
1367       SmallVector<uint32_t, 8> Idxs(DstNumElts);
1368       for (unsigned i = 0; i != SrcNumElts; ++i)
1369         Idxs[i] = i;
1370       for (unsigned i = SrcNumElts; i != DstNumElts; ++i)
1371         Idxs[i] = SrcNumElts;
1372       Rep = Builder.CreateShuffleVector(Op1, UndefV, Idxs);
1373 
1374       // Insert the second operand into the first operand.
1375 
1376       // Note that there is no guarantee that instruction lowering will actually
1377       // produce a vinsertf128 instruction for the created shuffles. In
1378       // particular, the 0 immediate case involves no lane changes, so it can
1379       // be handled as a blend.
1380 
1381       // Example of shuffle mask for 32-bit elements:
1382       // Imm = 1  <i32 0, i32 1, i32 2,  i32 3,  i32 8, i32 9, i32 10, i32 11>
1383       // Imm = 0  <i32 8, i32 9, i32 10, i32 11, i32 4, i32 5, i32 6,  i32 7 >
1384 
1385       // First fill with identify mask.
1386       for (unsigned i = 0; i != DstNumElts; ++i)
1387         Idxs[i] = i;
1388       // Then replace the elements where we need to insert.
1389       for (unsigned i = 0; i != SrcNumElts; ++i)
1390         Idxs[i + Imm * SrcNumElts] = i + DstNumElts;
1391       Rep = Builder.CreateShuffleVector(Op0, Rep, Idxs);
1392 
1393       // If the intrinsic has a mask operand, handle that.
1394       if (CI->getNumArgOperands() == 5)
1395         Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1396                             CI->getArgOperand(3));
1397     } else if (IsX86 && (Name.startswith("avx.vextractf128.") ||
1398                          Name == "avx2.vextracti128" ||
1399                          Name.startswith("avx512.mask.vextract"))) {
1400       Value *Op0 = CI->getArgOperand(0);
1401       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1402       unsigned DstNumElts = CI->getType()->getVectorNumElements();
1403       unsigned SrcNumElts = Op0->getType()->getVectorNumElements();
1404       unsigned Scale = SrcNumElts / DstNumElts;
1405 
1406       // Mask off the high bits of the immediate value; hardware ignores those.
1407       Imm = Imm % Scale;
1408 
1409       // Get indexes for the subvector of the input vector.
1410       SmallVector<uint32_t, 8> Idxs(DstNumElts);
1411       for (unsigned i = 0; i != DstNumElts; ++i) {
1412         Idxs[i] = i + (Imm * DstNumElts);
1413       }
1414       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1415 
1416       // If the intrinsic has a mask operand, handle that.
1417       if (CI->getNumArgOperands() == 4)
1418         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1419                             CI->getArgOperand(2));
1420     } else if (!IsX86 && Name == "stackprotectorcheck") {
1421       Rep = nullptr;
1422     } else if (IsX86 && (Name.startswith("avx512.mask.perm.df.") ||
1423                          Name.startswith("avx512.mask.perm.di."))) {
1424       Value *Op0 = CI->getArgOperand(0);
1425       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1426       VectorType *VecTy = cast<VectorType>(CI->getType());
1427       unsigned NumElts = VecTy->getNumElements();
1428 
1429       SmallVector<uint32_t, 8> Idxs(NumElts);
1430       for (unsigned i = 0; i != NumElts; ++i)
1431         Idxs[i] = (i & ~0x3) + ((Imm >> (2 * (i & 0x3))) & 3);
1432 
1433       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1434 
1435       if (CI->getNumArgOperands() == 4)
1436         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1437                             CI->getArgOperand(2));
1438     } else if (IsX86 && (Name.startswith("avx.vperm2f128.") ||
1439                          Name == "avx2.vperm2i128")) {
1440       // The immediate permute control byte looks like this:
1441       //    [1:0] - select 128 bits from sources for low half of destination
1442       //    [2]   - ignore
1443       //    [3]   - zero low half of destination
1444       //    [5:4] - select 128 bits from sources for high half of destination
1445       //    [6]   - ignore
1446       //    [7]   - zero high half of destination
1447 
1448       uint8_t Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1449 
1450       unsigned NumElts = CI->getType()->getVectorNumElements();
1451       unsigned HalfSize = NumElts / 2;
1452       SmallVector<uint32_t, 8> ShuffleMask(NumElts);
1453 
1454       // Determine which operand(s) are actually in use for this instruction.
1455       Value *V0 = (Imm & 0x02) ? CI->getArgOperand(1) : CI->getArgOperand(0);
1456       Value *V1 = (Imm & 0x20) ? CI->getArgOperand(1) : CI->getArgOperand(0);
1457 
1458       // If needed, replace operands based on zero mask.
1459       V0 = (Imm & 0x08) ? ConstantAggregateZero::get(CI->getType()) : V0;
1460       V1 = (Imm & 0x80) ? ConstantAggregateZero::get(CI->getType()) : V1;
1461 
1462       // Permute low half of result.
1463       unsigned StartIndex = (Imm & 0x01) ? HalfSize : 0;
1464       for (unsigned i = 0; i < HalfSize; ++i)
1465         ShuffleMask[i] = StartIndex + i;
1466 
1467       // Permute high half of result.
1468       StartIndex = (Imm & 0x10) ? HalfSize : 0;
1469       for (unsigned i = 0; i < HalfSize; ++i)
1470         ShuffleMask[i + HalfSize] = NumElts + StartIndex + i;
1471 
1472       Rep = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
1473 
1474     } else if (IsX86 && (Name.startswith("avx.vpermil.") ||
1475                          Name == "sse2.pshuf.d" ||
1476                          Name.startswith("avx512.mask.vpermil.p") ||
1477                          Name.startswith("avx512.mask.pshuf.d."))) {
1478       Value *Op0 = CI->getArgOperand(0);
1479       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1480       VectorType *VecTy = cast<VectorType>(CI->getType());
1481       unsigned NumElts = VecTy->getNumElements();
1482       // Calculate the size of each index in the immediate.
1483       unsigned IdxSize = 64 / VecTy->getScalarSizeInBits();
1484       unsigned IdxMask = ((1 << IdxSize) - 1);
1485 
1486       SmallVector<uint32_t, 8> Idxs(NumElts);
1487       // Lookup the bits for this element, wrapping around the immediate every
1488       // 8-bits. Elements are grouped into sets of 2 or 4 elements so we need
1489       // to offset by the first index of each group.
1490       for (unsigned i = 0; i != NumElts; ++i)
1491         Idxs[i] = ((Imm >> ((i * IdxSize) % 8)) & IdxMask) | (i & ~IdxMask);
1492 
1493       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1494 
1495       if (CI->getNumArgOperands() == 4)
1496         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1497                             CI->getArgOperand(2));
1498     } else if (IsX86 && (Name == "sse2.pshufl.w" ||
1499                          Name.startswith("avx512.mask.pshufl.w."))) {
1500       Value *Op0 = CI->getArgOperand(0);
1501       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1502       unsigned NumElts = CI->getType()->getVectorNumElements();
1503 
1504       SmallVector<uint32_t, 16> Idxs(NumElts);
1505       for (unsigned l = 0; l != NumElts; l += 8) {
1506         for (unsigned i = 0; i != 4; ++i)
1507           Idxs[i + l] = ((Imm >> (2 * i)) & 0x3) + l;
1508         for (unsigned i = 4; i != 8; ++i)
1509           Idxs[i + l] = i + l;
1510       }
1511 
1512       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1513 
1514       if (CI->getNumArgOperands() == 4)
1515         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1516                             CI->getArgOperand(2));
1517     } else if (IsX86 && (Name == "sse2.pshufh.w" ||
1518                          Name.startswith("avx512.mask.pshufh.w."))) {
1519       Value *Op0 = CI->getArgOperand(0);
1520       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(1))->getZExtValue();
1521       unsigned NumElts = CI->getType()->getVectorNumElements();
1522 
1523       SmallVector<uint32_t, 16> Idxs(NumElts);
1524       for (unsigned l = 0; l != NumElts; l += 8) {
1525         for (unsigned i = 0; i != 4; ++i)
1526           Idxs[i + l] = i + l;
1527         for (unsigned i = 0; i != 4; ++i)
1528           Idxs[i + l + 4] = ((Imm >> (2 * i)) & 0x3) + 4 + l;
1529       }
1530 
1531       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1532 
1533       if (CI->getNumArgOperands() == 4)
1534         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1535                             CI->getArgOperand(2));
1536     } else if (IsX86 && Name.startswith("avx512.mask.shuf.p")) {
1537       Value *Op0 = CI->getArgOperand(0);
1538       Value *Op1 = CI->getArgOperand(1);
1539       unsigned Imm = cast<ConstantInt>(CI->getArgOperand(2))->getZExtValue();
1540       unsigned NumElts = CI->getType()->getVectorNumElements();
1541 
1542       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1543       unsigned HalfLaneElts = NumLaneElts / 2;
1544 
1545       SmallVector<uint32_t, 16> Idxs(NumElts);
1546       for (unsigned i = 0; i != NumElts; ++i) {
1547         // Base index is the starting element of the lane.
1548         Idxs[i] = i - (i % NumLaneElts);
1549         // If we are half way through the lane switch to the other source.
1550         if ((i % NumLaneElts) >= HalfLaneElts)
1551           Idxs[i] += NumElts;
1552         // Now select the specific element. By adding HalfLaneElts bits from
1553         // the immediate. Wrapping around the immediate every 8-bits.
1554         Idxs[i] += (Imm >> ((i * HalfLaneElts) % 8)) & ((1 << HalfLaneElts) - 1);
1555       }
1556 
1557       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1558 
1559       Rep = EmitX86Select(Builder, CI->getArgOperand(4), Rep,
1560                           CI->getArgOperand(3));
1561     } else if (IsX86 && (Name.startswith("avx512.mask.movddup") ||
1562                          Name.startswith("avx512.mask.movshdup") ||
1563                          Name.startswith("avx512.mask.movsldup"))) {
1564       Value *Op0 = CI->getArgOperand(0);
1565       unsigned NumElts = CI->getType()->getVectorNumElements();
1566       unsigned NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1567 
1568       unsigned Offset = 0;
1569       if (Name.startswith("avx512.mask.movshdup."))
1570         Offset = 1;
1571 
1572       SmallVector<uint32_t, 16> Idxs(NumElts);
1573       for (unsigned l = 0; l != NumElts; l += NumLaneElts)
1574         for (unsigned i = 0; i != NumLaneElts; i += 2) {
1575           Idxs[i + l + 0] = i + l + Offset;
1576           Idxs[i + l + 1] = i + l + Offset;
1577         }
1578 
1579       Rep = Builder.CreateShuffleVector(Op0, Op0, Idxs);
1580 
1581       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1582                           CI->getArgOperand(1));
1583     } else if (IsX86 && (Name.startswith("avx512.mask.punpckl") ||
1584                          Name.startswith("avx512.mask.unpckl."))) {
1585       Value *Op0 = CI->getArgOperand(0);
1586       Value *Op1 = CI->getArgOperand(1);
1587       int NumElts = CI->getType()->getVectorNumElements();
1588       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1589 
1590       SmallVector<uint32_t, 64> Idxs(NumElts);
1591       for (int l = 0; l != NumElts; l += NumLaneElts)
1592         for (int i = 0; i != NumLaneElts; ++i)
1593           Idxs[i + l] = l + (i / 2) + NumElts * (i % 2);
1594 
1595       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1596 
1597       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1598                           CI->getArgOperand(2));
1599     } else if (IsX86 && (Name.startswith("avx512.mask.punpckh") ||
1600                          Name.startswith("avx512.mask.unpckh."))) {
1601       Value *Op0 = CI->getArgOperand(0);
1602       Value *Op1 = CI->getArgOperand(1);
1603       int NumElts = CI->getType()->getVectorNumElements();
1604       int NumLaneElts = 128/CI->getType()->getScalarSizeInBits();
1605 
1606       SmallVector<uint32_t, 64> Idxs(NumElts);
1607       for (int l = 0; l != NumElts; l += NumLaneElts)
1608         for (int i = 0; i != NumLaneElts; ++i)
1609           Idxs[i + l] = (NumLaneElts / 2) + l + (i / 2) + NumElts * (i % 2);
1610 
1611       Rep = Builder.CreateShuffleVector(Op0, Op1, Idxs);
1612 
1613       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1614                           CI->getArgOperand(2));
1615     } else if (IsX86 && Name.startswith("avx512.mask.pand.")) {
1616       Rep = Builder.CreateAnd(CI->getArgOperand(0), CI->getArgOperand(1));
1617       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1618                           CI->getArgOperand(2));
1619     } else if (IsX86 && Name.startswith("avx512.mask.pandn.")) {
1620       Rep = Builder.CreateAnd(Builder.CreateNot(CI->getArgOperand(0)),
1621                               CI->getArgOperand(1));
1622       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1623                           CI->getArgOperand(2));
1624     } else if (IsX86 && Name.startswith("avx512.mask.por.")) {
1625       Rep = Builder.CreateOr(CI->getArgOperand(0), CI->getArgOperand(1));
1626       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1627                           CI->getArgOperand(2));
1628     } else if (IsX86 && Name.startswith("avx512.mask.pxor.")) {
1629       Rep = Builder.CreateXor(CI->getArgOperand(0), CI->getArgOperand(1));
1630       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1631                           CI->getArgOperand(2));
1632     } else if (IsX86 && Name.startswith("avx512.mask.and.")) {
1633       VectorType *FTy = cast<VectorType>(CI->getType());
1634       VectorType *ITy = VectorType::getInteger(FTy);
1635       Rep = Builder.CreateAnd(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1636                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1637       Rep = Builder.CreateBitCast(Rep, FTy);
1638       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1639                           CI->getArgOperand(2));
1640     } else if (IsX86 && Name.startswith("avx512.mask.andn.")) {
1641       VectorType *FTy = cast<VectorType>(CI->getType());
1642       VectorType *ITy = VectorType::getInteger(FTy);
1643       Rep = Builder.CreateNot(Builder.CreateBitCast(CI->getArgOperand(0), ITy));
1644       Rep = Builder.CreateAnd(Rep,
1645                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1646       Rep = Builder.CreateBitCast(Rep, FTy);
1647       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1648                           CI->getArgOperand(2));
1649     } else if (IsX86 && Name.startswith("avx512.mask.or.")) {
1650       VectorType *FTy = cast<VectorType>(CI->getType());
1651       VectorType *ITy = VectorType::getInteger(FTy);
1652       Rep = Builder.CreateOr(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1653                              Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1654       Rep = Builder.CreateBitCast(Rep, FTy);
1655       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1656                           CI->getArgOperand(2));
1657     } else if (IsX86 && Name.startswith("avx512.mask.xor.")) {
1658       VectorType *FTy = cast<VectorType>(CI->getType());
1659       VectorType *ITy = VectorType::getInteger(FTy);
1660       Rep = Builder.CreateXor(Builder.CreateBitCast(CI->getArgOperand(0), ITy),
1661                               Builder.CreateBitCast(CI->getArgOperand(1), ITy));
1662       Rep = Builder.CreateBitCast(Rep, FTy);
1663       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1664                           CI->getArgOperand(2));
1665     } else if (IsX86 && Name.startswith("avx512.mask.padd.")) {
1666       Rep = Builder.CreateAdd(CI->getArgOperand(0), CI->getArgOperand(1));
1667       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1668                           CI->getArgOperand(2));
1669     } else if (IsX86 && Name.startswith("avx512.mask.psub.")) {
1670       Rep = Builder.CreateSub(CI->getArgOperand(0), CI->getArgOperand(1));
1671       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1672                           CI->getArgOperand(2));
1673     } else if (IsX86 && Name.startswith("avx512.mask.pmull.")) {
1674       Rep = Builder.CreateMul(CI->getArgOperand(0), CI->getArgOperand(1));
1675       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1676                           CI->getArgOperand(2));
1677     } else if (IsX86 && (Name.startswith("avx512.mask.add.p"))) {
1678       Rep = Builder.CreateFAdd(CI->getArgOperand(0), CI->getArgOperand(1));
1679       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1680                           CI->getArgOperand(2));
1681     } else if (IsX86 && Name.startswith("avx512.mask.div.p")) {
1682       Rep = Builder.CreateFDiv(CI->getArgOperand(0), CI->getArgOperand(1));
1683       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1684                           CI->getArgOperand(2));
1685     } else if (IsX86 && Name.startswith("avx512.mask.mul.p")) {
1686       Rep = Builder.CreateFMul(CI->getArgOperand(0), CI->getArgOperand(1));
1687       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1688                           CI->getArgOperand(2));
1689     } else if (IsX86 && Name.startswith("avx512.mask.sub.p")) {
1690       Rep = Builder.CreateFSub(CI->getArgOperand(0), CI->getArgOperand(1));
1691       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1692                           CI->getArgOperand(2));
1693     } else if (IsX86 && Name.startswith("avx512.mask.lzcnt.")) {
1694       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(),
1695                                                          Intrinsic::ctlz,
1696                                                          CI->getType()),
1697                                { CI->getArgOperand(0), Builder.getInt1(false) });
1698       Rep = EmitX86Select(Builder, CI->getArgOperand(2), Rep,
1699                           CI->getArgOperand(1));
1700     } else if (IsX86 && (Name.startswith("avx512.mask.max.p") ||
1701                          Name.startswith("avx512.mask.min.p"))) {
1702       bool IsMin = Name[13] == 'i';
1703       VectorType *VecTy = cast<VectorType>(CI->getType());
1704       unsigned VecWidth = VecTy->getPrimitiveSizeInBits();
1705       unsigned EltWidth = VecTy->getScalarSizeInBits();
1706       Intrinsic::ID IID;
1707       if (!IsMin && VecWidth == 128 && EltWidth == 32)
1708         IID = Intrinsic::x86_sse_max_ps;
1709       else if (!IsMin && VecWidth == 128 && EltWidth == 64)
1710         IID = Intrinsic::x86_sse2_max_pd;
1711       else if (!IsMin && VecWidth == 256 && EltWidth == 32)
1712         IID = Intrinsic::x86_avx_max_ps_256;
1713       else if (!IsMin && VecWidth == 256 && EltWidth == 64)
1714         IID = Intrinsic::x86_avx_max_pd_256;
1715       else if (IsMin && VecWidth == 128 && EltWidth == 32)
1716         IID = Intrinsic::x86_sse_min_ps;
1717       else if (IsMin && VecWidth == 128 && EltWidth == 64)
1718         IID = Intrinsic::x86_sse2_min_pd;
1719       else if (IsMin && VecWidth == 256 && EltWidth == 32)
1720         IID = Intrinsic::x86_avx_min_ps_256;
1721       else if (IsMin && VecWidth == 256 && EltWidth == 64)
1722         IID = Intrinsic::x86_avx_min_pd_256;
1723       else
1724         llvm_unreachable("Unexpected intrinsic");
1725 
1726       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1727                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1728       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1729                           CI->getArgOperand(2));
1730     } else if (IsX86 && Name.startswith("avx512.mask.pshuf.b.")) {
1731       VectorType *VecTy = cast<VectorType>(CI->getType());
1732       Intrinsic::ID IID;
1733       if (VecTy->getPrimitiveSizeInBits() == 128)
1734         IID = Intrinsic::x86_ssse3_pshuf_b_128;
1735       else if (VecTy->getPrimitiveSizeInBits() == 256)
1736         IID = Intrinsic::x86_avx2_pshuf_b;
1737       else if (VecTy->getPrimitiveSizeInBits() == 512)
1738         IID = Intrinsic::x86_avx512_pshuf_b_512;
1739       else
1740         llvm_unreachable("Unexpected intrinsic");
1741 
1742       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1743                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1744       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1745                           CI->getArgOperand(2));
1746     } else if (IsX86 && (Name.startswith("avx512.mask.pmul.dq.") ||
1747                          Name.startswith("avx512.mask.pmulu.dq."))) {
1748       bool IsUnsigned = Name[16] == 'u';
1749       VectorType *VecTy = cast<VectorType>(CI->getType());
1750       Intrinsic::ID IID;
1751       if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1752         IID = Intrinsic::x86_sse41_pmuldq;
1753       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1754         IID = Intrinsic::x86_avx2_pmul_dq;
1755       else if (!IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1756         IID = Intrinsic::x86_avx512_pmul_dq_512;
1757       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 128)
1758         IID = Intrinsic::x86_sse2_pmulu_dq;
1759       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 256)
1760         IID = Intrinsic::x86_avx2_pmulu_dq;
1761       else if (IsUnsigned && VecTy->getPrimitiveSizeInBits() == 512)
1762         IID = Intrinsic::x86_avx512_pmulu_dq_512;
1763       else
1764         llvm_unreachable("Unexpected intrinsic");
1765 
1766       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1767                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1768       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1769                           CI->getArgOperand(2));
1770     } else if (IsX86 && Name.startswith("avx512.mask.pack")) {
1771       bool IsUnsigned = Name[16] == 'u';
1772       bool IsDW = Name[18] == 'd';
1773       VectorType *VecTy = cast<VectorType>(CI->getType());
1774       Intrinsic::ID IID;
1775       if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1776         IID = Intrinsic::x86_sse2_packsswb_128;
1777       else if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1778         IID = Intrinsic::x86_avx2_packsswb;
1779       else if (!IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1780         IID = Intrinsic::x86_avx512_packsswb_512;
1781       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1782         IID = Intrinsic::x86_sse2_packssdw_128;
1783       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1784         IID = Intrinsic::x86_avx2_packssdw;
1785       else if (!IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1786         IID = Intrinsic::x86_avx512_packssdw_512;
1787       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1788         IID = Intrinsic::x86_sse2_packuswb_128;
1789       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1790         IID = Intrinsic::x86_avx2_packuswb;
1791       else if (IsUnsigned && !IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1792         IID = Intrinsic::x86_avx512_packuswb_512;
1793       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 128)
1794         IID = Intrinsic::x86_sse41_packusdw;
1795       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 256)
1796         IID = Intrinsic::x86_avx2_packusdw;
1797       else if (IsUnsigned && IsDW && VecTy->getPrimitiveSizeInBits() == 512)
1798         IID = Intrinsic::x86_avx512_packusdw_512;
1799       else
1800         llvm_unreachable("Unexpected intrinsic");
1801 
1802       Rep = Builder.CreateCall(Intrinsic::getDeclaration(F->getParent(), IID),
1803                                { CI->getArgOperand(0), CI->getArgOperand(1) });
1804       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
1805                           CI->getArgOperand(2));
1806     } else if (IsX86 && Name.startswith("avx512.mask.psll")) {
1807       bool IsImmediate = Name[16] == 'i' ||
1808                          (Name.size() > 18 && Name[18] == 'i');
1809       bool IsVariable = Name[16] == 'v';
1810       char Size = Name[16] == '.' ? Name[17] :
1811                   Name[17] == '.' ? Name[18] :
1812                   Name[18] == '.' ? Name[19] :
1813                                     Name[20];
1814 
1815       Intrinsic::ID IID;
1816       if (IsVariable && Name[17] != '.') {
1817         if (Size == 'd' && Name[17] == '2') // avx512.mask.psllv2.di
1818           IID = Intrinsic::x86_avx2_psllv_q;
1819         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psllv4.di
1820           IID = Intrinsic::x86_avx2_psllv_q_256;
1821         else if (Size == 's' && Name[17] == '4') // avx512.mask.psllv4.si
1822           IID = Intrinsic::x86_avx2_psllv_d;
1823         else if (Size == 's' && Name[17] == '8') // avx512.mask.psllv8.si
1824           IID = Intrinsic::x86_avx2_psllv_d_256;
1825         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psllv8.hi
1826           IID = Intrinsic::x86_avx512_psllv_w_128;
1827         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psllv16.hi
1828           IID = Intrinsic::x86_avx512_psllv_w_256;
1829         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psllv32hi
1830           IID = Intrinsic::x86_avx512_psllv_w_512;
1831         else
1832           llvm_unreachable("Unexpected size");
1833       } else if (Name.endswith(".128")) {
1834         if (Size == 'd') // avx512.mask.psll.d.128, avx512.mask.psll.di.128
1835           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_d
1836                             : Intrinsic::x86_sse2_psll_d;
1837         else if (Size == 'q') // avx512.mask.psll.q.128, avx512.mask.psll.qi.128
1838           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_q
1839                             : Intrinsic::x86_sse2_psll_q;
1840         else if (Size == 'w') // avx512.mask.psll.w.128, avx512.mask.psll.wi.128
1841           IID = IsImmediate ? Intrinsic::x86_sse2_pslli_w
1842                             : Intrinsic::x86_sse2_psll_w;
1843         else
1844           llvm_unreachable("Unexpected size");
1845       } else if (Name.endswith(".256")) {
1846         if (Size == 'd') // avx512.mask.psll.d.256, avx512.mask.psll.di.256
1847           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_d
1848                             : Intrinsic::x86_avx2_psll_d;
1849         else if (Size == 'q') // avx512.mask.psll.q.256, avx512.mask.psll.qi.256
1850           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_q
1851                             : Intrinsic::x86_avx2_psll_q;
1852         else if (Size == 'w') // avx512.mask.psll.w.256, avx512.mask.psll.wi.256
1853           IID = IsImmediate ? Intrinsic::x86_avx2_pslli_w
1854                             : Intrinsic::x86_avx2_psll_w;
1855         else
1856           llvm_unreachable("Unexpected size");
1857       } else {
1858         if (Size == 'd') // psll.di.512, pslli.d, psll.d, psllv.d.512
1859           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_d_512 :
1860                 IsVariable  ? Intrinsic::x86_avx512_psllv_d_512 :
1861                               Intrinsic::x86_avx512_psll_d_512;
1862         else if (Size == 'q') // psll.qi.512, pslli.q, psll.q, psllv.q.512
1863           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_q_512 :
1864                 IsVariable  ? Intrinsic::x86_avx512_psllv_q_512 :
1865                               Intrinsic::x86_avx512_psll_q_512;
1866         else if (Size == 'w') // psll.wi.512, pslli.w, psll.w
1867           IID = IsImmediate ? Intrinsic::x86_avx512_pslli_w_512
1868                             : Intrinsic::x86_avx512_psll_w_512;
1869         else
1870           llvm_unreachable("Unexpected size");
1871       }
1872 
1873       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
1874     } else if (IsX86 && Name.startswith("avx512.mask.psrl")) {
1875       bool IsImmediate = Name[16] == 'i' ||
1876                          (Name.size() > 18 && Name[18] == 'i');
1877       bool IsVariable = Name[16] == 'v';
1878       char Size = Name[16] == '.' ? Name[17] :
1879                   Name[17] == '.' ? Name[18] :
1880                   Name[18] == '.' ? Name[19] :
1881                                     Name[20];
1882 
1883       Intrinsic::ID IID;
1884       if (IsVariable && Name[17] != '.') {
1885         if (Size == 'd' && Name[17] == '2') // avx512.mask.psrlv2.di
1886           IID = Intrinsic::x86_avx2_psrlv_q;
1887         else if (Size == 'd' && Name[17] == '4') // avx512.mask.psrlv4.di
1888           IID = Intrinsic::x86_avx2_psrlv_q_256;
1889         else if (Size == 's' && Name[17] == '4') // avx512.mask.psrlv4.si
1890           IID = Intrinsic::x86_avx2_psrlv_d;
1891         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrlv8.si
1892           IID = Intrinsic::x86_avx2_psrlv_d_256;
1893         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrlv8.hi
1894           IID = Intrinsic::x86_avx512_psrlv_w_128;
1895         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrlv16.hi
1896           IID = Intrinsic::x86_avx512_psrlv_w_256;
1897         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrlv32hi
1898           IID = Intrinsic::x86_avx512_psrlv_w_512;
1899         else
1900           llvm_unreachable("Unexpected size");
1901       } else if (Name.endswith(".128")) {
1902         if (Size == 'd') // avx512.mask.psrl.d.128, avx512.mask.psrl.di.128
1903           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_d
1904                             : Intrinsic::x86_sse2_psrl_d;
1905         else if (Size == 'q') // avx512.mask.psrl.q.128, avx512.mask.psrl.qi.128
1906           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_q
1907                             : Intrinsic::x86_sse2_psrl_q;
1908         else if (Size == 'w') // avx512.mask.psrl.w.128, avx512.mask.psrl.wi.128
1909           IID = IsImmediate ? Intrinsic::x86_sse2_psrli_w
1910                             : Intrinsic::x86_sse2_psrl_w;
1911         else
1912           llvm_unreachable("Unexpected size");
1913       } else if (Name.endswith(".256")) {
1914         if (Size == 'd') // avx512.mask.psrl.d.256, avx512.mask.psrl.di.256
1915           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_d
1916                             : Intrinsic::x86_avx2_psrl_d;
1917         else if (Size == 'q') // avx512.mask.psrl.q.256, avx512.mask.psrl.qi.256
1918           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_q
1919                             : Intrinsic::x86_avx2_psrl_q;
1920         else if (Size == 'w') // avx512.mask.psrl.w.256, avx512.mask.psrl.wi.256
1921           IID = IsImmediate ? Intrinsic::x86_avx2_psrli_w
1922                             : Intrinsic::x86_avx2_psrl_w;
1923         else
1924           llvm_unreachable("Unexpected size");
1925       } else {
1926         if (Size == 'd') // psrl.di.512, psrli.d, psrl.d, psrl.d.512
1927           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_d_512 :
1928                 IsVariable  ? Intrinsic::x86_avx512_psrlv_d_512 :
1929                               Intrinsic::x86_avx512_psrl_d_512;
1930         else if (Size == 'q') // psrl.qi.512, psrli.q, psrl.q, psrl.q.512
1931           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_q_512 :
1932                 IsVariable  ? Intrinsic::x86_avx512_psrlv_q_512 :
1933                               Intrinsic::x86_avx512_psrl_q_512;
1934         else if (Size == 'w') // psrl.wi.512, psrli.w, psrl.w)
1935           IID = IsImmediate ? Intrinsic::x86_avx512_psrli_w_512
1936                             : Intrinsic::x86_avx512_psrl_w_512;
1937         else
1938           llvm_unreachable("Unexpected size");
1939       }
1940 
1941       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
1942     } else if (IsX86 && Name.startswith("avx512.mask.psra")) {
1943       bool IsImmediate = Name[16] == 'i' ||
1944                          (Name.size() > 18 && Name[18] == 'i');
1945       bool IsVariable = Name[16] == 'v';
1946       char Size = Name[16] == '.' ? Name[17] :
1947                   Name[17] == '.' ? Name[18] :
1948                   Name[18] == '.' ? Name[19] :
1949                                     Name[20];
1950 
1951       Intrinsic::ID IID;
1952       if (IsVariable && Name[17] != '.') {
1953         if (Size == 's' && Name[17] == '4') // avx512.mask.psrav4.si
1954           IID = Intrinsic::x86_avx2_psrav_d;
1955         else if (Size == 's' && Name[17] == '8') // avx512.mask.psrav8.si
1956           IID = Intrinsic::x86_avx2_psrav_d_256;
1957         else if (Size == 'h' && Name[17] == '8') // avx512.mask.psrav8.hi
1958           IID = Intrinsic::x86_avx512_psrav_w_128;
1959         else if (Size == 'h' && Name[17] == '1') // avx512.mask.psrav16.hi
1960           IID = Intrinsic::x86_avx512_psrav_w_256;
1961         else if (Name[17] == '3' && Name[18] == '2') // avx512.mask.psrav32hi
1962           IID = Intrinsic::x86_avx512_psrav_w_512;
1963         else
1964           llvm_unreachable("Unexpected size");
1965       } else if (Name.endswith(".128")) {
1966         if (Size == 'd') // avx512.mask.psra.d.128, avx512.mask.psra.di.128
1967           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_d
1968                             : Intrinsic::x86_sse2_psra_d;
1969         else if (Size == 'q') // avx512.mask.psra.q.128, avx512.mask.psra.qi.128
1970           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_128 :
1971                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_128 :
1972                               Intrinsic::x86_avx512_psra_q_128;
1973         else if (Size == 'w') // avx512.mask.psra.w.128, avx512.mask.psra.wi.128
1974           IID = IsImmediate ? Intrinsic::x86_sse2_psrai_w
1975                             : Intrinsic::x86_sse2_psra_w;
1976         else
1977           llvm_unreachable("Unexpected size");
1978       } else if (Name.endswith(".256")) {
1979         if (Size == 'd') // avx512.mask.psra.d.256, avx512.mask.psra.di.256
1980           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_d
1981                             : Intrinsic::x86_avx2_psra_d;
1982         else if (Size == 'q') // avx512.mask.psra.q.256, avx512.mask.psra.qi.256
1983           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_256 :
1984                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_256 :
1985                               Intrinsic::x86_avx512_psra_q_256;
1986         else if (Size == 'w') // avx512.mask.psra.w.256, avx512.mask.psra.wi.256
1987           IID = IsImmediate ? Intrinsic::x86_avx2_psrai_w
1988                             : Intrinsic::x86_avx2_psra_w;
1989         else
1990           llvm_unreachable("Unexpected size");
1991       } else {
1992         if (Size == 'd') // psra.di.512, psrai.d, psra.d, psrav.d.512
1993           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_d_512 :
1994                 IsVariable  ? Intrinsic::x86_avx512_psrav_d_512 :
1995                               Intrinsic::x86_avx512_psra_d_512;
1996         else if (Size == 'q') // psra.qi.512, psrai.q, psra.q
1997           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_q_512 :
1998                 IsVariable  ? Intrinsic::x86_avx512_psrav_q_512 :
1999                               Intrinsic::x86_avx512_psra_q_512;
2000         else if (Size == 'w') // psra.wi.512, psrai.w, psra.w
2001           IID = IsImmediate ? Intrinsic::x86_avx512_psrai_w_512
2002                             : Intrinsic::x86_avx512_psra_w_512;
2003         else
2004           llvm_unreachable("Unexpected size");
2005       }
2006 
2007       Rep = UpgradeX86MaskedShift(Builder, *CI, IID);
2008     } else if (IsX86 && Name.startswith("avx512.mask.move.s")) {
2009       Rep = upgradeMaskedMove(Builder, *CI);
2010     } else if (IsX86 && Name.startswith("avx512.cvtmask2")) {
2011       Rep = UpgradeMaskToInt(Builder, *CI);
2012     } else if (IsX86 && Name.startswith("avx512.mask.vpermilvar.")) {
2013       Intrinsic::ID IID;
2014       if (Name.endswith("ps.128"))
2015         IID = Intrinsic::x86_avx_vpermilvar_ps;
2016       else if (Name.endswith("pd.128"))
2017         IID = Intrinsic::x86_avx_vpermilvar_pd;
2018       else if (Name.endswith("ps.256"))
2019         IID = Intrinsic::x86_avx_vpermilvar_ps_256;
2020       else if (Name.endswith("pd.256"))
2021         IID = Intrinsic::x86_avx_vpermilvar_pd_256;
2022       else if (Name.endswith("ps.512"))
2023         IID = Intrinsic::x86_avx512_vpermilvar_ps_512;
2024       else if (Name.endswith("pd.512"))
2025         IID = Intrinsic::x86_avx512_vpermilvar_pd_512;
2026       else
2027         llvm_unreachable("Unexpected vpermilvar intrinsic");
2028 
2029       Function *Intrin = Intrinsic::getDeclaration(F->getParent(), IID);
2030       Rep = Builder.CreateCall(Intrin,
2031                                { CI->getArgOperand(0), CI->getArgOperand(1) });
2032       Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2033                           CI->getArgOperand(2));
2034     } else if (IsX86 && Name.endswith(".movntdqa")) {
2035       Module *M = F->getParent();
2036       MDNode *Node = MDNode::get(
2037           C, ConstantAsMetadata::get(ConstantInt::get(Type::getInt32Ty(C), 1)));
2038 
2039       Value *Ptr = CI->getArgOperand(0);
2040       VectorType *VTy = cast<VectorType>(CI->getType());
2041 
2042       // Convert the type of the pointer to a pointer to the stored type.
2043       Value *BC =
2044           Builder.CreateBitCast(Ptr, PointerType::getUnqual(VTy), "cast");
2045       LoadInst *LI = Builder.CreateAlignedLoad(BC, VTy->getBitWidth() / 8);
2046       LI->setMetadata(M->getMDKindID("nontemporal"), Node);
2047       Rep = LI;
2048     } else if (IsX86 &&
2049                (Name.startswith("sse2.pavg") || Name.startswith("avx2.pavg") ||
2050                 Name.startswith("avx512.mask.pavg"))) {
2051       // llvm.x86.sse2.pavg.b/w, llvm.x86.avx2.pavg.b/w,
2052       // llvm.x86.avx512.mask.pavg.b/w
2053       Value *A = CI->getArgOperand(0);
2054       Value *B = CI->getArgOperand(1);
2055       VectorType *ZextType = VectorType::getExtendedElementVectorType(
2056           cast<VectorType>(A->getType()));
2057       Value *ExtendedA = Builder.CreateZExt(A, ZextType);
2058       Value *ExtendedB = Builder.CreateZExt(B, ZextType);
2059       Value *Sum = Builder.CreateAdd(ExtendedA, ExtendedB);
2060       Value *AddOne = Builder.CreateAdd(Sum, ConstantInt::get(ZextType, 1));
2061       Value *ShiftR = Builder.CreateLShr(AddOne, ConstantInt::get(ZextType, 1));
2062       Rep = Builder.CreateTrunc(ShiftR, A->getType());
2063       if (CI->getNumArgOperands() > 2) {
2064         Rep = EmitX86Select(Builder, CI->getArgOperand(3), Rep,
2065                             CI->getArgOperand(2));
2066       }
2067     } else if (IsNVVM && (Name == "abs.i" || Name == "abs.ll")) {
2068       Value *Arg = CI->getArgOperand(0);
2069       Value *Neg = Builder.CreateNeg(Arg, "neg");
2070       Value *Cmp = Builder.CreateICmpSGE(
2071           Arg, llvm::Constant::getNullValue(Arg->getType()), "abs.cond");
2072       Rep = Builder.CreateSelect(Cmp, Arg, Neg, "abs");
2073     } else if (IsNVVM && (Name == "max.i" || Name == "max.ll" ||
2074                           Name == "max.ui" || Name == "max.ull")) {
2075       Value *Arg0 = CI->getArgOperand(0);
2076       Value *Arg1 = CI->getArgOperand(1);
2077       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
2078                        ? Builder.CreateICmpUGE(Arg0, Arg1, "max.cond")
2079                        : Builder.CreateICmpSGE(Arg0, Arg1, "max.cond");
2080       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "max");
2081     } else if (IsNVVM && (Name == "min.i" || Name == "min.ll" ||
2082                           Name == "min.ui" || Name == "min.ull")) {
2083       Value *Arg0 = CI->getArgOperand(0);
2084       Value *Arg1 = CI->getArgOperand(1);
2085       Value *Cmp = Name.endswith(".ui") || Name.endswith(".ull")
2086                        ? Builder.CreateICmpULE(Arg0, Arg1, "min.cond")
2087                        : Builder.CreateICmpSLE(Arg0, Arg1, "min.cond");
2088       Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "min");
2089     } else if (IsNVVM && Name == "clz.ll") {
2090       // llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 and returns an i64.
2091       Value *Arg = CI->getArgOperand(0);
2092       Value *Ctlz = Builder.CreateCall(
2093           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctlz,
2094                                     {Arg->getType()}),
2095           {Arg, Builder.getFalse()}, "ctlz");
2096       Rep = Builder.CreateTrunc(Ctlz, Builder.getInt32Ty(), "ctlz.trunc");
2097     } else if (IsNVVM && Name == "popc.ll") {
2098       // llvm.nvvm.popc.ll returns an i32, but llvm.ctpop.i64 and returns an
2099       // i64.
2100       Value *Arg = CI->getArgOperand(0);
2101       Value *Popc = Builder.CreateCall(
2102           Intrinsic::getDeclaration(F->getParent(), Intrinsic::ctpop,
2103                                     {Arg->getType()}),
2104           Arg, "ctpop");
2105       Rep = Builder.CreateTrunc(Popc, Builder.getInt32Ty(), "ctpop.trunc");
2106     } else if (IsNVVM && Name == "h2f") {
2107       Rep = Builder.CreateCall(Intrinsic::getDeclaration(
2108                                    F->getParent(), Intrinsic::convert_from_fp16,
2109                                    {Builder.getFloatTy()}),
2110                                CI->getArgOperand(0), "h2f");
2111     } else {
2112       llvm_unreachable("Unknown function for CallInst upgrade.");
2113     }
2114 
2115     if (Rep)
2116       CI->replaceAllUsesWith(Rep);
2117     CI->eraseFromParent();
2118     return;
2119   }
2120 
2121   CallInst *NewCall = nullptr;
2122   switch (NewFn->getIntrinsicID()) {
2123   default: {
2124     // Handle generic mangling change, but nothing else
2125     assert(
2126         (CI->getCalledFunction()->getName() != NewFn->getName()) &&
2127         "Unknown function for CallInst upgrade and isn't just a name change");
2128     CI->setCalledFunction(NewFn);
2129     return;
2130   }
2131 
2132   case Intrinsic::arm_neon_vld1:
2133   case Intrinsic::arm_neon_vld2:
2134   case Intrinsic::arm_neon_vld3:
2135   case Intrinsic::arm_neon_vld4:
2136   case Intrinsic::arm_neon_vld2lane:
2137   case Intrinsic::arm_neon_vld3lane:
2138   case Intrinsic::arm_neon_vld4lane:
2139   case Intrinsic::arm_neon_vst1:
2140   case Intrinsic::arm_neon_vst2:
2141   case Intrinsic::arm_neon_vst3:
2142   case Intrinsic::arm_neon_vst4:
2143   case Intrinsic::arm_neon_vst2lane:
2144   case Intrinsic::arm_neon_vst3lane:
2145   case Intrinsic::arm_neon_vst4lane: {
2146     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2147                                  CI->arg_operands().end());
2148     NewCall = Builder.CreateCall(NewFn, Args);
2149     break;
2150   }
2151 
2152   case Intrinsic::bitreverse:
2153     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2154     break;
2155 
2156   case Intrinsic::ctlz:
2157   case Intrinsic::cttz:
2158     assert(CI->getNumArgOperands() == 1 &&
2159            "Mismatch between function args and call args");
2160     NewCall =
2161         Builder.CreateCall(NewFn, {CI->getArgOperand(0), Builder.getFalse()});
2162     break;
2163 
2164   case Intrinsic::objectsize: {
2165     Value *NullIsUnknownSize = CI->getNumArgOperands() == 2
2166                                    ? Builder.getFalse()
2167                                    : CI->getArgOperand(2);
2168     NewCall = Builder.CreateCall(
2169         NewFn, {CI->getArgOperand(0), CI->getArgOperand(1), NullIsUnknownSize});
2170     break;
2171   }
2172 
2173   case Intrinsic::ctpop:
2174     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2175     break;
2176 
2177   case Intrinsic::convert_from_fp16:
2178     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(0)});
2179     break;
2180 
2181   case Intrinsic::dbg_value:
2182     // Upgrade from the old version that had an extra offset argument.
2183     assert(CI->getNumArgOperands() == 4);
2184     // Drop nonzero offsets instead of attempting to upgrade them.
2185     if (auto *Offset = dyn_cast_or_null<Constant>(CI->getArgOperand(1)))
2186       if (Offset->isZeroValue()) {
2187         NewCall = Builder.CreateCall(
2188             NewFn,
2189             {CI->getArgOperand(0), CI->getArgOperand(2), CI->getArgOperand(3)});
2190         break;
2191       }
2192     CI->eraseFromParent();
2193     return;
2194 
2195   case Intrinsic::x86_xop_vfrcz_ss:
2196   case Intrinsic::x86_xop_vfrcz_sd:
2197     NewCall = Builder.CreateCall(NewFn, {CI->getArgOperand(1)});
2198     break;
2199 
2200   case Intrinsic::x86_xop_vpermil2pd:
2201   case Intrinsic::x86_xop_vpermil2ps:
2202   case Intrinsic::x86_xop_vpermil2pd_256:
2203   case Intrinsic::x86_xop_vpermil2ps_256: {
2204     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2205                                  CI->arg_operands().end());
2206     VectorType *FltIdxTy = cast<VectorType>(Args[2]->getType());
2207     VectorType *IntIdxTy = VectorType::getInteger(FltIdxTy);
2208     Args[2] = Builder.CreateBitCast(Args[2], IntIdxTy);
2209     NewCall = Builder.CreateCall(NewFn, Args);
2210     break;
2211   }
2212 
2213   case Intrinsic::x86_sse41_ptestc:
2214   case Intrinsic::x86_sse41_ptestz:
2215   case Intrinsic::x86_sse41_ptestnzc: {
2216     // The arguments for these intrinsics used to be v4f32, and changed
2217     // to v2i64. This is purely a nop, since those are bitwise intrinsics.
2218     // So, the only thing required is a bitcast for both arguments.
2219     // First, check the arguments have the old type.
2220     Value *Arg0 = CI->getArgOperand(0);
2221     if (Arg0->getType() != VectorType::get(Type::getFloatTy(C), 4))
2222       return;
2223 
2224     // Old intrinsic, add bitcasts
2225     Value *Arg1 = CI->getArgOperand(1);
2226 
2227     Type *NewVecTy = VectorType::get(Type::getInt64Ty(C), 2);
2228 
2229     Value *BC0 = Builder.CreateBitCast(Arg0, NewVecTy, "cast");
2230     Value *BC1 = Builder.CreateBitCast(Arg1, NewVecTy, "cast");
2231 
2232     NewCall = Builder.CreateCall(NewFn, {BC0, BC1});
2233     break;
2234   }
2235 
2236   case Intrinsic::x86_sse41_insertps:
2237   case Intrinsic::x86_sse41_dppd:
2238   case Intrinsic::x86_sse41_dpps:
2239   case Intrinsic::x86_sse41_mpsadbw:
2240   case Intrinsic::x86_avx_dp_ps_256:
2241   case Intrinsic::x86_avx2_mpsadbw: {
2242     // Need to truncate the last argument from i32 to i8 -- this argument models
2243     // an inherently 8-bit immediate operand to these x86 instructions.
2244     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2245                                  CI->arg_operands().end());
2246 
2247     // Replace the last argument with a trunc.
2248     Args.back() = Builder.CreateTrunc(Args.back(), Type::getInt8Ty(C), "trunc");
2249     NewCall = Builder.CreateCall(NewFn, Args);
2250     break;
2251   }
2252 
2253   case Intrinsic::thread_pointer: {
2254     NewCall = Builder.CreateCall(NewFn, {});
2255     break;
2256   }
2257 
2258   case Intrinsic::invariant_start:
2259   case Intrinsic::invariant_end:
2260   case Intrinsic::masked_load:
2261   case Intrinsic::masked_store:
2262   case Intrinsic::masked_gather:
2263   case Intrinsic::masked_scatter: {
2264     SmallVector<Value *, 4> Args(CI->arg_operands().begin(),
2265                                  CI->arg_operands().end());
2266     NewCall = Builder.CreateCall(NewFn, Args);
2267     break;
2268   }
2269   }
2270   assert(NewCall && "Should have either set this variable or returned through "
2271                     "the default case");
2272   std::string Name = CI->getName();
2273   if (!Name.empty()) {
2274     CI->setName(Name + ".old");
2275     NewCall->setName(Name);
2276   }
2277   CI->replaceAllUsesWith(NewCall);
2278   CI->eraseFromParent();
2279 }
2280 
2281 void llvm::UpgradeCallsToIntrinsic(Function *F) {
2282   assert(F && "Illegal attempt to upgrade a non-existent intrinsic.");
2283 
2284   // Check if this function should be upgraded and get the replacement function
2285   // if there is one.
2286   Function *NewFn;
2287   if (UpgradeIntrinsicFunction(F, NewFn)) {
2288     // Replace all users of the old function with the new function or new
2289     // instructions. This is not a range loop because the call is deleted.
2290     for (auto UI = F->user_begin(), UE = F->user_end(); UI != UE; )
2291       if (CallInst *CI = dyn_cast<CallInst>(*UI++))
2292         UpgradeIntrinsicCall(CI, NewFn);
2293 
2294     // Remove old function, no longer used, from the module.
2295     F->eraseFromParent();
2296   }
2297 }
2298 
2299 MDNode *llvm::UpgradeTBAANode(MDNode &MD) {
2300   // Check if the tag uses struct-path aware TBAA format.
2301   if (isa<MDNode>(MD.getOperand(0)) && MD.getNumOperands() >= 3)
2302     return &MD;
2303 
2304   auto &Context = MD.getContext();
2305   if (MD.getNumOperands() == 3) {
2306     Metadata *Elts[] = {MD.getOperand(0), MD.getOperand(1)};
2307     MDNode *ScalarType = MDNode::get(Context, Elts);
2308     // Create a MDNode <ScalarType, ScalarType, offset 0, const>
2309     Metadata *Elts2[] = {ScalarType, ScalarType,
2310                          ConstantAsMetadata::get(
2311                              Constant::getNullValue(Type::getInt64Ty(Context))),
2312                          MD.getOperand(2)};
2313     return MDNode::get(Context, Elts2);
2314   }
2315   // Create a MDNode <MD, MD, offset 0>
2316   Metadata *Elts[] = {&MD, &MD, ConstantAsMetadata::get(Constant::getNullValue(
2317                                     Type::getInt64Ty(Context)))};
2318   return MDNode::get(Context, Elts);
2319 }
2320 
2321 Instruction *llvm::UpgradeBitCastInst(unsigned Opc, Value *V, Type *DestTy,
2322                                       Instruction *&Temp) {
2323   if (Opc != Instruction::BitCast)
2324     return nullptr;
2325 
2326   Temp = nullptr;
2327   Type *SrcTy = V->getType();
2328   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
2329       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
2330     LLVMContext &Context = V->getContext();
2331 
2332     // We have no information about target data layout, so we assume that
2333     // the maximum pointer size is 64bit.
2334     Type *MidTy = Type::getInt64Ty(Context);
2335     Temp = CastInst::Create(Instruction::PtrToInt, V, MidTy);
2336 
2337     return CastInst::Create(Instruction::IntToPtr, Temp, DestTy);
2338   }
2339 
2340   return nullptr;
2341 }
2342 
2343 Value *llvm::UpgradeBitCastExpr(unsigned Opc, Constant *C, Type *DestTy) {
2344   if (Opc != Instruction::BitCast)
2345     return nullptr;
2346 
2347   Type *SrcTy = C->getType();
2348   if (SrcTy->isPtrOrPtrVectorTy() && DestTy->isPtrOrPtrVectorTy() &&
2349       SrcTy->getPointerAddressSpace() != DestTy->getPointerAddressSpace()) {
2350     LLVMContext &Context = C->getContext();
2351 
2352     // We have no information about target data layout, so we assume that
2353     // the maximum pointer size is 64bit.
2354     Type *MidTy = Type::getInt64Ty(Context);
2355 
2356     return ConstantExpr::getIntToPtr(ConstantExpr::getPtrToInt(C, MidTy),
2357                                      DestTy);
2358   }
2359 
2360   return nullptr;
2361 }
2362 
2363 /// Check the debug info version number, if it is out-dated, drop the debug
2364 /// info. Return true if module is modified.
2365 bool llvm::UpgradeDebugInfo(Module &M) {
2366   unsigned Version = getDebugMetadataVersionFromModule(M);
2367   if (Version == DEBUG_METADATA_VERSION)
2368     return false;
2369 
2370   bool RetCode = StripDebugInfo(M);
2371   if (RetCode) {
2372     DiagnosticInfoDebugMetadataVersion DiagVersion(M, Version);
2373     M.getContext().diagnose(DiagVersion);
2374   }
2375   return RetCode;
2376 }
2377 
2378 bool llvm::UpgradeModuleFlags(Module &M) {
2379   NamedMDNode *ModFlags = M.getModuleFlagsMetadata();
2380   if (!ModFlags)
2381     return false;
2382 
2383   bool HasObjCFlag = false, HasClassProperties = false, Changed = false;
2384   for (unsigned I = 0, E = ModFlags->getNumOperands(); I != E; ++I) {
2385     MDNode *Op = ModFlags->getOperand(I);
2386     if (Op->getNumOperands() != 3)
2387       continue;
2388     MDString *ID = dyn_cast_or_null<MDString>(Op->getOperand(1));
2389     if (!ID)
2390       continue;
2391     if (ID->getString() == "Objective-C Image Info Version")
2392       HasObjCFlag = true;
2393     if (ID->getString() == "Objective-C Class Properties")
2394       HasClassProperties = true;
2395     // Upgrade PIC/PIE Module Flags. The module flag behavior for these two
2396     // field was Error and now they are Max.
2397     if (ID->getString() == "PIC Level" || ID->getString() == "PIE Level") {
2398       if (auto *Behavior =
2399               mdconst::dyn_extract_or_null<ConstantInt>(Op->getOperand(0))) {
2400         if (Behavior->getLimitedValue() == Module::Error) {
2401           Type *Int32Ty = Type::getInt32Ty(M.getContext());
2402           Metadata *Ops[3] = {
2403               ConstantAsMetadata::get(ConstantInt::get(Int32Ty, Module::Max)),
2404               MDString::get(M.getContext(), ID->getString()),
2405               Op->getOperand(2)};
2406           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
2407           Changed = true;
2408         }
2409       }
2410     }
2411     // Upgrade Objective-C Image Info Section. Removed the whitespce in the
2412     // section name so that llvm-lto will not complain about mismatching
2413     // module flags that is functionally the same.
2414     if (ID->getString() == "Objective-C Image Info Section") {
2415       if (auto *Value = dyn_cast_or_null<MDString>(Op->getOperand(2))) {
2416         SmallVector<StringRef, 4> ValueComp;
2417         Value->getString().split(ValueComp, " ");
2418         if (ValueComp.size() != 1) {
2419           std::string NewValue;
2420           for (auto &S : ValueComp)
2421             NewValue += S.str();
2422           Metadata *Ops[3] = {Op->getOperand(0), Op->getOperand(1),
2423                               MDString::get(M.getContext(), NewValue)};
2424           ModFlags->setOperand(I, MDNode::get(M.getContext(), Ops));
2425           Changed = true;
2426         }
2427       }
2428     }
2429   }
2430 
2431   // "Objective-C Class Properties" is recently added for Objective-C. We
2432   // upgrade ObjC bitcodes to contain a "Objective-C Class Properties" module
2433   // flag of value 0, so we can correclty downgrade this flag when trying to
2434   // link an ObjC bitcode without this module flag with an ObjC bitcode with
2435   // this module flag.
2436   if (HasObjCFlag && !HasClassProperties) {
2437     M.addModuleFlag(llvm::Module::Override, "Objective-C Class Properties",
2438                     (uint32_t)0);
2439     Changed = true;
2440   }
2441 
2442   return Changed;
2443 }
2444 
2445 static bool isOldLoopArgument(Metadata *MD) {
2446   auto *T = dyn_cast_or_null<MDTuple>(MD);
2447   if (!T)
2448     return false;
2449   if (T->getNumOperands() < 1)
2450     return false;
2451   auto *S = dyn_cast_or_null<MDString>(T->getOperand(0));
2452   if (!S)
2453     return false;
2454   return S->getString().startswith("llvm.vectorizer.");
2455 }
2456 
2457 static MDString *upgradeLoopTag(LLVMContext &C, StringRef OldTag) {
2458   StringRef OldPrefix = "llvm.vectorizer.";
2459   assert(OldTag.startswith(OldPrefix) && "Expected old prefix");
2460 
2461   if (OldTag == "llvm.vectorizer.unroll")
2462     return MDString::get(C, "llvm.loop.interleave.count");
2463 
2464   return MDString::get(
2465       C, (Twine("llvm.loop.vectorize.") + OldTag.drop_front(OldPrefix.size()))
2466              .str());
2467 }
2468 
2469 static Metadata *upgradeLoopArgument(Metadata *MD) {
2470   auto *T = dyn_cast_or_null<MDTuple>(MD);
2471   if (!T)
2472     return MD;
2473   if (T->getNumOperands() < 1)
2474     return MD;
2475   auto *OldTag = dyn_cast_or_null<MDString>(T->getOperand(0));
2476   if (!OldTag)
2477     return MD;
2478   if (!OldTag->getString().startswith("llvm.vectorizer."))
2479     return MD;
2480 
2481   // This has an old tag.  Upgrade it.
2482   SmallVector<Metadata *, 8> Ops;
2483   Ops.reserve(T->getNumOperands());
2484   Ops.push_back(upgradeLoopTag(T->getContext(), OldTag->getString()));
2485   for (unsigned I = 1, E = T->getNumOperands(); I != E; ++I)
2486     Ops.push_back(T->getOperand(I));
2487 
2488   return MDTuple::get(T->getContext(), Ops);
2489 }
2490 
2491 MDNode *llvm::upgradeInstructionLoopAttachment(MDNode &N) {
2492   auto *T = dyn_cast<MDTuple>(&N);
2493   if (!T)
2494     return &N;
2495 
2496   if (none_of(T->operands(), isOldLoopArgument))
2497     return &N;
2498 
2499   SmallVector<Metadata *, 8> Ops;
2500   Ops.reserve(T->getNumOperands());
2501   for (Metadata *MD : T->operands())
2502     Ops.push_back(upgradeLoopArgument(MD));
2503 
2504   return MDTuple::get(T->getContext(), Ops);
2505 }
2506